Log In Sign Up

Layout Generation and Completion with Self-attention

by   Kamal Gupta, et al.

We address the problem of layout generation for diverse domains such as images, documents, and mobile applications. A layout is a set of graphical elements, belonging to one or more categories, placed together in a meaningful way. Generating a new layout or extending an existing layout requires understanding the relationships between these graphical elements. To do this, we propose a novel framework, LayoutTransformer, that leverages a self-attention based approach to learn contextual relationships between layout elements and generate layouts in a given domain. The proposed model improves upon the state-of-the-art approaches in layout generation in four ways. First, our model can generate a new layout either from an empty set or add more elements to a partial layout starting from an initial set of elements. Second, as the approach is attention-based, we can visualize which previous elements the model is attending to predict the next element, thereby providing an interpretable sequence of layout elements. Third, our model can easily scale to support both a large number of element categories and a large number of elements per layout. Finally, the model also produces an embedding for various element categories, which can be used to explore the relationships between the categories. We demonstrate with experiments that our model can produce meaningful layouts in diverse settings such as object bounding boxes in scenes (COCO bounding boxes), documents (PubLayNet), and mobile applications (RICO dataset).


page 2

page 8

page 13

page 20

page 21

page 22

page 23


Generative Layout Modeling using Constraint Graphs

We propose a new generative model for layout generation. We generate lay...

LayoutGAN: Generating Graphic Layouts with Wireframe Discriminators

Layout is important for graphic design and scene generation. We propose ...

LayoutBERT: Masked Language Layout Model for Object Insertion

Image compositing is one of the most fundamental steps in creative workf...

Geometry Aligned Variational Transformer for Image-conditioned Layout Generation

Layout generation is a novel task in computer vision, which combines the...

READ: Recursive Autoencoders for Document Layout Generation

Layout is a fundamental component of any graphic design. Creating large ...

Bib2vec: An Embedding-based Search System for Bibliographic Information

We propose a novel embedding model that represents relationships among s...

Reverse-Engineering Information Presentations: Recovering Hierarchical Grouping from Layouts of Visual Elements

Visual elements in an information presentation are often spatially and s...

Code Repositories

1 Introduction

In the real world, there exists a strong relationship between different objects that are found in the same environment [36, 34]. For example, a dining table usually has chairs around it; a surfboard is found near the sea; horses do not ride cars; etc.  Biederman [2] provided strong evidence in cognitive neuroscience that perceiving and understanding a scene involves two related processes: perception and comprehension. Perception deals with processing the visual signal or the appearance of a scene. Comprehension deals with understanding the schema of a scene [2], where this schema (or layout) can be characterized by contextual relationships between objects (e.g., support, occlusion, and relative likelihood, position, and size [2]). For generative models that synthesize scenes, this evidence underpins the importance of two factors that contribute to the realism or plausibility of a generated scene: layout, i.e.the arrangement of different objects, and their appearance (in terms of pixels). Therefore, generating a realistic scene necessitates both these factors to be plausible.

Figure 1: (a) LayoutTransformer can generate multiple layouts consisting of variable number of elements starting from an empty canvas. (b) We can use tools such as Layout2Im [48] to generate image from layout (best viewed in color)

The advancements in the generative models for image synthesis have primarily targeted plausibility of the appearance signal by generating incredibly realistic images of objects (e.g., faces [20, 19], animals [3, 46]). In order to generate complex scenes [6, 15, 17, 26, 32, 47, 48], most methods require proxy representations for layouts to be provided as inputs (e.g.scene segmentation [16, 40], textual descriptions [26, 47, 33], scene graphs [17]). We argue that to plausibly generate large and complex scenes without such proxies, it is necessary to understand and generate the layout of the scene, in terms of contextual relationships between various objects present in the scene.

The layout of a scene, capturing what objects occupy what parts of the scene, is an incredibly rich representation. A plausible layout needs to follow prior knowledge of world regularities [2, 7]; for example, the sky is above the sea, indoor scenes have certain furniture arrangements etc.. Learning to generate layouts is a challenging problem due to the variability of real-world layouts. Each scene contains a small fraction of possible objects, objects can be present in a wide range of locations, the number of objects varies for each scene and so do the contextual relationships between objects (e.g., a person can carry a surfboard or ride a surfboard, a person can ride horse but not carry it). We parameterize each object or element of the layout by semantic attributes/categories, location, and size. In order to realize plausible semantic relationships, a generative model for layouts should be able to look at all existing objects and propose placement of a new object.

We propose a sequential and iterative approach for modeling layouts that uses self-attention to look at existing layout elements. Our generative process can start from an empty set or an unordered set of elements already present in the scene, and can iteratively attend to existing elements to generate a new element. Moreover, by predicting either to stop or to generate the next element, our sequential approach can generate variable length layouts.

Our approach can be readily plugged into many scene generation frameworks (e.g., Layout2Image [48], GauGAN [31]). However, layout generation is not limited to scene layouts. Several stand-alone applications require generating layouts or templates. For instance, in the UI design of mobile apps and websites [8, 28], an automated model for generating plausible layouts can significantly decrease the manual effort and cost of building such apps and websites. Finally, a model to create layouts can potentially help generate synthetic data for augmentation tasks [44, 4] or 3D scenes [5, 43, 42].

We summarize the contributions of our work as follows:

  • We develop a simple yet powerful auto-regressive model for generating layouts that can synthesize new layouts, complete partial layouts, and compute likelihood of existing layouts. Our self-attention approach allows us to visualize what existing elements are important for generating the next category in the sequence,

  • We propose modeling position and size of layout elements with discrete multinomial distribution that enables our approach to generalize across very diverse data distributions,

  • We present an exciting finding – encouraging a model to understand layouts results in feature representations that capture the semantic relationships between objects automatically (without explicitly using semantic embeddings, like word2vec [29]). This demonstrates the utility of the task of layout generation as a proxy-task for learning semantic representations,

  • We show the performance and adaptability of our model on four layout datasets: MNIST Layout [25], Rico Mobile App Wireframes [8], PubLayNet Documents [45], and COCO Bounding Boxes [27]

2 Related Work

Generative models. Deep generative models in recent years have shown a great promise in terms of faithfully learning a given data distribution and sampling from it. Approaches such as variational auto-encoders [22] try to maximize approximate log-likelihood of data generated from a known distribution. Auto-regressive and flow-based approaches such as Pixel-RNN [37], RealNVP [9] can compute exact log-likelihood but are inefficient to sample from. GANs [11] are arguably the most popular generative models demonstrating state-of-the-art image generation results [3, 20], but do not allow log-likelihood computation. Most of these approaches and their variations work well when generating entire images, especially for datasets of images with a single entity or object. While these models can generate realistic objects, they often fail to capture global semantic and geometric relations between objects, which are needed to generate more complex realistic scenes [26].

Scene generation. Most works in 2D or 3D scene synthesis generate a scene conditioned on a sentence [26, 47, 33], a scene graph [17, 24, 1], a layout [10, 13, 16, 41] or an existing image [23]. These involved pipelines are often trained and evaluated end-to-end, and surprisingly little work has been done to evaluate the layout generation component itself. Given the input, some works generate a fixed layout and diverse scenes [48], while other works generate diverse layouts and scenes [17, 26]. Again the focus of all these works is on the quality of the final image and not the feasibility of the layout itself. In this work, we evaluate the layout modeling capabilities of two of the recent works [17, 26] that have layout generation as an intermediate step.

Layout generation. Synthesising scene layouts is a relatively under-explored problem, mainly because generative models do not work well in practice when modeling sets of elements rather than images. LayoutGAN [25] attempts to solve the problem by starting with maximum number of possible elements in the scene and modifying their geometric and semantic properties. LayoutVAE [18] starts with a label set, i.e., categories of all the elements present in the layout and then generates a feasible layout of the scene. Wang et al.  [40, 39] generate layout of an indoor room starting from top-down view of the room. However, their method is very specific to indoor rooms data and make assumptions about presence of walls, roof etc., and hence cannot be easily extended to other datasets. Zheng et al. [49] attempt to generate document layouts given the images, keywords and category of the document.

Our approach, LayoutTransformer, offers several advantages over current layout generation approaches without sacrificing their benefits. Unlike [25, 49] autoregressive nature of model allows us to generate layouts of arbitrary lengths as well as start with partial layouts. We observe that modeling the position and size of layout elements as discrete values (as discussed in 3.1) helps us realize better performance on datasets, such as documents and app wireframes, where bounding boxes of layouts are typically axis aligned. Finally, assumptions of various kind of inputs limit the applicability of existing methods to diverse datasets. For example, to use scene graphs [17, 24] as input, relationships need to be redefined for different datasets, text [26, 47, 33] need not exist for documents or wireframes, and to images can be used as input for documents but not for images. We get rid of lot of assumptions and simplify our layout generation pipeline to such an extent that it can be used to synthesize layouts for very diverse datasets. With thorough experimentation, we show the superiority of LayoutTransformer over three diverse real world datasets COCO Bounding Boxes [27], PubLayNet Documents [45], and Rico Mobile App Wireframes [8].

3 LayoutTransformer

Figure 2: The architecture for the Transformer model depicted for a toy example. It takes layout elements as input and predicts next layout elements as output. During training, we use teacher forcing, i.e.

, use the groundtruth layout tokens as input to a multi-head decoder block. The first layer of this block is a masked self-attention layer, which allows the model to see only the previous elements in order to predict the current element. We pad each layout with a special

token in the beginning and token in the end. To generate new layouts, we perform beam search starting with just the token or a partial sequence.

Layouts are distinct from scenes or images in several ways. There is a strong non-local relationships between different elements in layouts. For example, presence of a small bird in corner of a scene changes distribution of objects present in rest of the scene, or a figure in document decides where the text can go. In case of images, on the other hand, local relationships play a more dominant role, i.e., pixels close to each other are likely to be similar in values. While CNN based architectures such as VAE and GANs are excellent at generating pixels or images, an intuitive way to go about modeling distribution of complex scenes would be to use an auto-regressive model that can look at all existing elements, near or far, to generate semantic relationships followed by a convolutional architecture to generate the final image.

In this section, we propose LayoutTransformer, an auto-regressive self-attention network architecture to model the layouts. It allows us to learn non-local semantic relationships between layout elements and also give us flexibility to work with variable length layouts. We first describe the problem setup and follow it up with details of network architecture and training.

3.1 Problem Setup

We represent each layout as a sequence of graphical elements comprising the layout. For two-dimensional datasets such as documents, images, and wireframes, each graphical element can be further represented by a bounding box of category , located at , of size

. The goal of the layout model is now to learn the joint distribution of category, location, and size of layout elements represented by a tuple

. We order all the elements in the raster scan order i.e.first by coordinate, followed by coordinate. After reordering, we concatenate all the graphical elements in a flat sequence. We also append two special symbols and to denote start and end of sequence. Hence, our layout of graphical elements can be now represented as a sequence

For simplicity, we use to represent any element in the flattened sequence of the tuples, i.e.,

. We now pose the problem of modeling this joint distribution as product over series of conditional distributions using chain rule:

Representation of layout element. As introduced earlier, each layout element is represented by its category and the bounding box enclosing it. Instead of treating the location and size of bounding boxes as continuous variables, we model them as a discrete distribution where each

is obtained as the output of a softmax layer. Apart from allowing us to represent each of

in a simple and consistent manner, this strategy has the additional advantage that it does not assume a prior on position and size of bounding boxes and lets the network model them arbitrarily. If we divide our layout in grid cells, each element of

can be represented by a one-hot encoded vector of size

, where we use to denote the size of vocabulary

In Fig. 5, we show that discretizing position and shape in this way is particularly advantageous for datasets such as document layouts when bounding boxes are aligned to each other. We also discuss a variation of this strategy in ablation studies.

Ordering of elements. The sequence of elements is important in order to train our model. We use a simple raster scan order of layout elements in our case. We show the impact of removing this ordering strategy and using an arbitrary order in ablation studies.

3.2 Network Architecture and Training

We use a Transformer Decoder [38]

to estimate the above joint probability distribution. Fig. 

2 shows the network architecture for a toy example. Each layout element gets mapped to a -dimensional embedding such that . These embeddings are then passed to a sequence of masked self-attention layers. The masking is done so that model predicts the probability of an element using only the embeddings of previous layout elements. This means

where represent the masked multi-headed self-attention transformer decoder module. It includes a softmax layer in the end to normalize the output values between 0 and 1. Instead of using a standard cross-entropy loss, we follow the approach commonly used in Transformer-like architectures. For every groundtruth element in the sequence , we create a pseudo groundtruth distribution using Label Smoothing [30] with high confidence at the groundtruth index and rest of the mass distributed uniformly. We then minimize the KL-Divergence loss of predictions with this distribution using . Label Smoothing impacts the perplexity of the model adversely but prevents the model from becoming overconfident.

In our base model, we use , , and in the decoder. Label smoothing uses an . We observe that our model is quite robust to these choices, as we show in the ablation studies. The rest of the details of network architecture and training are in the supplementary material.

3.3 Sampling layouts

At training and validation time, we use teacher forcing, i.e., since we know all the sequences, we use groundtruth sequences of variable lengths to train our model efficiently. To sample a new layout, we can start off with either just a start of sequence token or a set of tokens . A naïve decoding strategy would be greedy decoding, i.e., we predict the next element with the highest probability, append it to the existing sequence, and repeat till we reach the end of the sequence . A better way would be to do a beam search [35], i.e., keep track of most likely sequences while decoding to generate multiple possible layouts starting from the same initial sequence.

4 Experiments

In this section, we first discuss datasets used for evaluation, followed by qualitative results of our approach on these datasets. We then analyze the performance of LayoutTransformer on general and dataset-specific quantitative metrics.

Figure 3: Generated layouts. First column (in each dataset) shows a random layout from validation data rendered on a blank image. We use part of the sequence in this layout (from validation data) to generate the most probable layout using our model. Second column onwards, we show initial layout given to LayoutTransformer for completion followed by layout as completed by LayoutTransformer. We skip the label names in case of RICO dataset for clarity.

4.1 Datasets

We evaluate the performance of our approach on multiple diverse datasets. Specifically, we use a toy MNIST Layout dataset as proposed in LayoutGAN [25], Rico Mobile App Wireframes [8, 28], COCO bounding boxes [27] and PubLayNet Documents [45]. Note that each of the datasets involves a pre-processing step, and we tried to faithfully replicate these steps as provided in original publications [25, 18]. We will release the code for pre-processing, our approach, and experiments for reproducibility and future reference.

MNIST Layout. Following LayoutGAN [25], in all images, we consider pixels with values greater than a fixed threshold (fixed to in all experiments) as foreground pixels. Each image is now represented by a set of randomly selected foreground pixel indices with a minimum of and a maximum of indices per image. Overall, we get training and validation layouts from the original train/val split of MNIST. The mean and median length of layouts is and , respectively.

COCO bounding boxes. COCO bounding boxes dataset is obtained using bounding box annotations in COCO Panoptic 2017 dataset [27]. We ignore the images where the isCrowd flag is true following the LayoutVAE [18] approach. The bounding boxes come from all thing and stuff categories. Our final dataset has layouts from COCO train split with a median length of elements and layouts from COCO valid split with a median length of .

Rico Mobile App Wireframes. Rico mobile app dataset [8, 28] consists of layout information of more than unique UI screens from over android apps. Each layout consists of one or more of the categories of graphical elements such as text, image, icon etc.  A complete list of these elements and frequency of their appearances is provided in the supplementary material. Overall, we get 62951 layouts in Rico with a median length of 36.

PubLayNet. PubLayNet [45] is a large scale document dataset recently released by IBM. It consists of over million documents collected from PubMed Central. The layouts are annotated with element categories - text, title, list, label, and figure. We filter out the document layouts with over 128 elements. Our final dataset has layouts from PubLayNet train split with a median length of elements and layouts from PubLayNet dev split with a median length of .

4.2 Baseline Models

We consider following baseline approaches:

LayoutVAE. LayoutVAE [18] uses a similar representation for layout and consists of two separate autoregressive VAE models. Starting from a label set, which consists of categories of elements that will be present in a generated layout, their CountVAE generates counts of each of the elements of the label set. After that BoundingBoxVAE, generates the location and size of each occurrence of the bounding box.

ObjGAN. ObjGAN [26] provides an object-attention based GAN framework for text to image synthesis. An intermediate step in their image synthesis approach is to generate a bounding box layout given a sentence using a BiLSTM (trained independently). We adopt this step of the ObjGAN framework to our problem setup. Instead of sentences we provide categories of all layout elements as input to the ObjGAN and synthesize all the elements’ bounding boxes.

sg2im. Image generation from scene graph [17] attempts to generate complex scenes given scene graph of the image by first generating a layout of the scene using graph convolutions and then using the layout to generate complete scene using GANs. The system is trained in an end-to-end fashion. Since sg2im requires a scene graph input, following the approach of  [17], we create a scene graph from the input and reproduce the input layout using the scene graph.

4.3 Qualitative Evaluation

Generated layout samples. Figure 3 shows some of the generated samples of our model from different datasets. Note that our model can generate samples from empty or partial layouts. We demonstrate this by taking partial layouts from validation data and generating a full layout with greedy decoding.

Figure 4: Nearest neighbors from training data. Column 1 shows the initial layout provided (for completion) to LayoutTransformer. Column 2 shows the layout as generated by LayoutTransformer. Column 3, 4 and 5 show the 3 closest neighbors from training dataset. We use chamfer distance on bounding box coordinates to obtain the nearest neighbors from the dataset.

Nearest neighbors. To see if our model is memorizing the training dataset, we compute nearest neighbors of generated layouts using chamfer distance on top-left and bottom-right bounding box coordinates of layout elements. Figure 4 shows the nearest neighbors of some of the generated layouts from the training dataset. Our model is able to generate novel layouts not present in the training data.

Figure 5: Generated samples for PubLayNet dataset using LayoutVAE [18] and our method. Our method produces aligned bounding boxes for various synthesized layout elements such as figure, text, title and tables.
Figure 6: Visualizing attention. (a) Image source for the layout (b) In each row, the model is predicting one element at a time (shown in a green bounding box). While predicting that element, the model pays the most attention to previously predicted bounding boxes (in red). For example, in the first row, “snow” gets the highest attention score while predicting “skis”. Similarly in the last column, “skis” get the highest attention while predicting “person”.

Visualizing attention. The self-attention based approach proposed enables us to visualize which existing elements are being attending to while the model is generating a new element. This is demonstrated in Figure 6.

Figure 7: Word embeddings as learned by LayoutTransformer on COCO bounding boxes. Words are colored by their super-categories provided in the COCO dataset. We see that semantically similar categories cluster together, e.g., animals, fruits, furniture, etc. Cats and dogs are closer to each other compared to sheep, zebra, or cow. Also, semantically similar words from different super-categories (such as curtain, window-blind, and mirror) are close in LayoutTransformer’s embedding.
Figure 8: We plot the distribution of x- and y-coordinates of the center of bounding boxes (normalized between 0 and 1). The y-coordinate is more informative (e.g., sky is usually on the top of the image while road and sea are at the bottom). Distributions for generated layouts and real layouts tend to be similar.

4.4 Semantics Emerge via Layout

We posited earlier that capturing layout should capture contextual relationships between various elements. We provide further evidence of our argument in three ways. We visualize the 2D-tsne plot of the learned embeddings for categories, as shown in Figure 8. We observe that super-categories from COCO are clustered together in the embedding space of the model. Certain categories such as window-blind and curtain (which belong to different super-categories) also appear close to each other. Table 2 captures the most frequent bigrams and trigrams (categories that co-occur) in real and synthesized layouts. Table 2 shows word2vec [29] style analogies being captured by embeddings learned by our model. Note that the model was trained to generate layouts and we did not specify any additional objective function for analogical reasoning task. These observations are in line with observations of Gupta et al [12].

Real Ours Real Ours
other person other person person other person other person clothes
person other person clothes other person clothes person clothes tie
person clothes clothes tie person handbag person tree grass other
clothes person grass other person clothes person grass other person
chair person other dining table person chair person wall-concrete other person
person chair tree grass chair person chair grass other cow
sky-other tree wall-concrete other person other clothes tree other person
car person person other person backpack person person clothes person
person handbag sky-other tree person car person other dining table table
handbag person clothes person person skis person person other person
Table 2: Analogies. We demonstrate linguistic nuances being captured by our category embeddings by attempting to solve word2vec [29] style analogies.
Analogy Nearest neighbors
snowboard:snow::surfboard:? waterdrops, sea, sand
car:road::train:? railroad, platform, gravel
sky-other:clouds::playingfield:? net, cage, wall-panel
mouse:keyboard::spoon:? knife, fork, oven
fruit:table::flower:? potted plant, mirror-stuff
Table 1: Bigrams and trigrams. We consider the most frequent pairs and triplets of (distinct) categories in real vs.generated layouts.

4.5 Quantitative Evaluation

Quantitative evaluation methods of the layout generation problem differ from dataset to dataset. In this section, we discuss some of these methods applicable to the datasets under consideration.

Method IS Real 16.1 [18] + [48] 7.1 Ours + [48] 7.5 Method FID [18] + [48] 64.1 Ours + [48] 57
Table 3: Image generation from layouts - We use L2Im [48] to convert layouts to images. (Left) The first row shows images using layouts from validation data. The second row shows generated novel layouts converted to image using the same model. The third row shows layouts generated by LayoutVAE converted to image in a similar manner. (Right) We compute Inception Score (IS) and Fréchet Inception Distance (FID) to compare quality and diversity of generated images from our layouts as compared to real layouts.

Downstream Task - Image generation. To evaluate the ability of layout generation approaches in generating plausible layouts for the COCO dataset we use Layout2Im [48] to generate an image from a layout. Table 3 shows images generated from layouts by our method and LayoutVAE. We also compute compute Inception Score (IS) and Fréchet Inception Distance (FID) to compare quality and diversity of generated images. Our method improves upon LayoutVAE in both the metrics.

Negative log-likelihood. For each of the datasets, Table 5 shows the negative log-likelihood of all the layouts in validation set. The results indicate that our approach generates more plausible layouts (more details on this are provided in the supplementary material).

Dataset statistics. Depending on the dataset and definition of graphical elements, we can define statistics that layouts should follow. For Rico wireframes and PubLayNet docs, we compare two important statistics of layouts in Table 5. Overlap represents the intersection over union (IoU) of various layout elements. Generally in these datasets, elements do not overap with each other and Overlap is small. Coverage indicates the percentage of canvas covered by the layout elements. The table shows that layouts generated by our method resemble real data statistics better than LayoutGAN and LayoutVAE.

Model COCO Rico PubLayNet
sg2im [17] 6.24 7.43 7.12
ObjGAN [26] 5.24 4.21 4.20
LayoutVAE [18] 3.29 2.54 2.45
Ours 2.28 1.07 1.10
Table 5: Spatial distribution analysis for the RICO dataset. As we limit the resolution for location and size of the bounding boxes to , our model is to be compared to the lower resolution version of the real data. Closer the values to real data, better is the performance. Clearly, the statistics of our layouts are the more similar to the real data statistics than sg2im, ObjGAN and LayoutVAE. All values in the table are percentages (std in parenthesis)
Rico PubLayNet
Methods Coverage Overlap Coverage Overlap.
sg2im [17] 25.2 (46) 16.5 (31) 30.2 (26) 3.4 (12)
ObjGAN [26] 39.2 (33) 36.4 (29) 38.9 (12) 8.2 (7)
LayoutVAE [18] 41.5 (29) 34.1 (27) 40.1 (11) 14.5 (11)
Ours 33.6 (27) 23.7 (33) 47.0 (12) 0.13 (1.5)
Real Data () 30.2 (25) 20.5 (30) 47.8 (9) 0.02 (0.5)
Real Data 36.6 (27) 22.4 (32) 57.1 (10) 0.1 (0.6)
Table 4: Negative log-likelihood of all the layouts in validation set (lower the better). For each of the approach we compute log-likelihood using importance sampling as described in  [18]. For LayoutVAE, we use our own implementation (since official implementation is not provided
# params COCO Rico PubLayNet
19.2 2.28 1.07 1.10
19.1 1.69 0.98 0.88
19.2 1.97 1.03 0.95
19.3 2.67 1.26 1.28
19.6 3.12 1.44 1.46
Table 7: Effect of on NLL
d # params COCO Rico PubLayNet
512 19.2 2.28 1.07 1.10
32 0.8 2.51 1.56 1.26
64 1.7 2.43 1.40 1.19
128 3.6 2.37 1.29 1.57
256 8.1 2.32 1.20 1.56
Table 6: Effect of on NLL
# params COCO Rico PubLayNet
6 19.2 2.28 1.07 1.10
2 6.6 2.31 1.18 1.13
4 12.9 2.30 1.12 1.07
8 25.5 2.28 1.11 1.07
Table 9:

Effect of other hyperparameters on NLL

Order Split-XY Loss # params COCO Rico PubLayNet
raster Yes NLL 19.2 2.28 1.07 1.10
random 19.2 2.68 1.76 1.46
No 21.2 3.74 2.12 1.87
LS 19.2 1.96 0.88 0.88
Table 8: Effect of on NLL

4.6 Ablation studies

We evaluate the importance of different model components with negative log-likelihood on COCO layouts. The ablation studies clearly show the following:

Small, medium and large elements: NLL of our model for COCO large, medium, and small boxes is 2.4, 2.5, and 1.8 respectively. We observe that even though discretizing box coordinates introduces approximation errors, it later allows our model to be agnostic to large vs small objects.

Varying : decides the resolution of the layout. Increasing it allows us to generate finer layouts but at the expense of a model with more parameters. Also, as we increase the , NLL increases, suggesting that we might need to train the model with more data to get similar performance (Table 7).

Size of embedding: Increasing the size of the embedding improves the NLL, but at the cost of increased number of parameters (Table 7).

Model depth: Increasing the depth of the model , does not significantly improve the results (Table 9). We fix the in all our experiments.

Ordering of the elements: The self-attention layer in our model is invariant to the ordering of elements. Therefore, while predicting the next element of the layout, we do not consider the order in which the elements were added. However, in our experiments, we observed that predicting the elements in a simple raster scan order of their position improves the model performance both visually and in terms of negative log-likelihood. This is intuitive as filling the elements in a pre-defined order is an easier problem. We leave the task of optimal ordering of layout elements to generate layouts for future research (Table 9).

Discretization strategy: Instead of the next bounding box, we tried predicting the x-coordinate and y-coordinate of the bounding box together (refer to the Split-xy column of Table 9). This increases the vocabulary size of the model (since we use possible locations instead of ) and in turn the number of hyper-parameters with decline in model performance. An upside of this approach is that generating new layouts takes less time as we have to make half as many predictions for each element of the layout (Table 9).

Loss: We tried two different losses, label smoothing [30]

and NLL. Although optimizing using NLL gives better validation performance in terms of NLL (as is expected), we do not find much difference in the qualitative performance when using either loss function. (Table 


5 Conclusion

We propose LayoutTransformer, a novel approach to generate layouts of graphical elements. Our model uses self-attention model to capture contextual relationship between different layout elements and generate novel layouts. We show that our model is better than previously proposed models for layout generation due to its ability to synthesize layouts from an empty set or complete a partial layout. The model can also produce layouts with a variable number of elements and categories. We show that our model can produce qualitatively better layouts than the state-of-the-art approaches for diverse datasets such as Rico Mobile App Wireframes, COCO bounding boxes, and PubLayNet documents. While we demonstrated results for our approach by generating layouts in two dimensions, this framework is applicable to three dimensional scenes as well. One limitation of the model is that while it is capable of predicting size and location of objects of the scene, it cannot be easily extended to predict object masks. We will explore these directions in future work.

Acknowledgements. We thank Yuting Zhang, Luis Goncalves, Stefano Soatto and Guha Balakrishnan for many helpful discussions. This work was partially supported by DARPA via ARO contract number W911NF2020009.


  • [1] O. Ashual and L. Wolf (2019) Specifying object attributes and relations in interactive scene generation. In

    Proceedings of the IEEE International Conference on Computer Vision

    pp. 4561–4569. Cited by: §2.
  • [2] I. Biederman (2017) On the semantics of a glance at a scene. In Perceptual organization, pp. 213–253. Cited by: §1, §1.
  • [3] A. Brock, J. Donahue, and K. Simonyan (2018) Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096. Cited by: §1, §2.
  • [4] S. Capobianco and S. Marinai (2017) DocEmul: a toolkit to generate structured historical documents. CoRR abs/1710.03474. External Links: Link, 1710.03474 Cited by: §1.
  • [5] A. X. Chang, W. Monroe, M. Savva, C. Potts, and C. D. Manning (2015) Text to 3d scene generation with rich lexical grounding. CoRR abs/1505.06289. External Links: Link, 1505.06289 Cited by: §1.
  • [6] Q. Chen and V. Koltun (2017) Photographic image synthesis with cascaded refinement networks. In Proceedings of the IEEE International Conference on Computer Vision, pp. 1511–1520. Cited by: §1.
  • [7] X. Chen, A. Shrivastava, and A. Gupta (2013) Neil: extracting visual knowledge from web data. In Proceedings of the IEEE International Conference on Computer Vision, pp. 1409–1416. Cited by: §1.
  • [8] B. Deka, Z. Huang, C. Franzen, J. Hibschman, D. Afergan, Y. Li, J. Nichols, and R. Kumar (2017) Rico: a mobile app dataset for building data-driven design applications. In Proceedings of the 30th Annual Symposium on User Interface Software and Technology, UIST ’17. Cited by: 4th item, §1, §2, §4.1, §4.1.
  • [9] L. Dinh, J. Sohl-Dickstein, and S. Bengio (2016) Density estimation using real nvp. arXiv preprint arXiv:1605.08803. Cited by: §2.
  • [10] H. Dong, S. Yu, C. Wu, and Y. Guo (2017) Semantic image synthesis via adversarial learning. In Proceedings of the IEEE International Conference on Computer Vision, pp. 5706–5714. Cited by: §2.
  • [11] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio (2014) Generative adversarial nets. In Advances in neural information processing systems, pp. 2672–2680. Cited by: §2.
  • [12] T. Gupta, A. Schwing, and D. Hoiem (2019) ViCo: word embeddings from visual co-occurrences. In Proceedings of the IEEE International Conference on Computer Vision, pp. 7425–7434. Cited by: §4.4.
  • [13] T. Hinz, S. Heinrich, and S. Wermter (2019) Generating multiple objects at spatially distinct locations. CoRR abs/1901.00686. External Links: Link, 1901.00686 Cited by: §2.
  • [14] A. Holtzman, J. Buys, M. Forbes, and Y. Choi (2019) The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751. Cited by: Figure 9.
  • [15] S. Hong, D. Yang, J. Choi, and H. Lee (2018) Inferring semantic layout for hierarchical text-to-image synthesis. In

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    pp. 7986–7994. Cited by: §1.
  • [16] P. Isola, J. Zhu, T. Zhou, and A. A. Efros (2016) Image-to-image translation with conditional adversarial networks. arxiv. Cited by: §1, §2.
  • [17] J. Johnson, A. Gupta, and L. Fei-Fei (2018) Image generation from scene graphs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1219–1228. Cited by: 10(c), 8(c), §1, §2, §2, §4.2, Table 5.
  • [18] A. A. Jyothi, T. Durand, J. He, L. Sigal, and G. Mori (2019) Layoutvae: stochastic scene layout generation from a label set. arXiv preprint arXiv:1907.10719. Cited by: 10(b), 8(b), 9(b), §0.B.1, §2, Figure 5, §4.1, §4.1, §4.2, Table 3, Table 5.
  • [19] T. Karras, T. Aila, S. Laine, and J. Lehtinen (2017) Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196. Cited by: §1.
  • [20] T. Karras, S. Laine, and T. Aila (2019) A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4401–4410. Cited by: §1, §2.
  • [21] D. P. Kingma and J. Ba (2014) Adam: a method for stochastic optimization. arXiv preprint arXiv:1412.6980. Cited by: Appendix 0.A.
  • [22] D. P. Kingma and M. Welling (2013) Auto-encoding variational bayes. External Links: 1312.6114 Cited by: §2.
  • [23] D. Lee, S. Liu, J. Gu, M. Liu, M. Yang, and J. Kautz (2018) Context-aware synthesis and placement of object instances. CoRR abs/1812.02350. External Links: Link, 1812.02350 Cited by: §2.
  • [24] B. Li, B. Zhuang, M. Li, and J. Gu (2019) Seq-sg2sl: inferring semantic layout from scene graph through sequence to sequence learning. In Proceedings of the IEEE International Conference on Computer Vision, pp. 7435–7443. Cited by: §2, §2.
  • [25] J. Li, J. Yang, A. Hertzmann, J. Zhang, and T. Xu (2019) LayoutGAN: generating graphic layouts with wireframe discriminators. arXiv preprint arXiv:1901.06767. Cited by: §0.B.4, §0.B.5, 4th item, §2, §2, §4.1, §4.1.
  • [26] W. Li, P. Zhang, L. Zhang, Q. Huang, X. He, S. Lyu, and J. Gao (2019) Object-driven text-to-image synthesis via adversarial training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 12174–12182. Cited by: 10(d), 8(d), §1, §2, §2, §2, §4.2, Table 5.
  • [27] T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick (2014) Microsoft coco: common objects in context. In European conference on computer vision, pp. 740–755. Cited by: 4th item, §2, §4.1, §4.1.
  • [28] T. F. Liu, M. Craft, J. Situ, E. Yumer, R. Mech, and R. Kumar (2018) Learning design semantics for mobile apps. In The 31st Annual ACM Symposium on User Interface Software and Technology, UIST ’18, New York, NY, USA, pp. 569–579. External Links: ISBN 978-1-4503-5948-1, Link, Document Cited by: §1, §4.1, §4.1.
  • [29] T. Mikolov, K. Chen, G. Corrado, and J. Dean (2013) Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781. Cited by: 3rd item, §4.4, Table 2.
  • [30] R. Müller, S. Kornblith, and G. E. Hinton (2019) When does label smoothing help?. CoRR abs/1906.02629. External Links: Link, 1906.02629 Cited by: §3.2, §4.6.
  • [31] T. Park, M. Liu, T. Wang, and J. Zhu (2019) GauGAN: semantic image synthesis with spatially adaptive normalization. In ACM SIGGRAPH 2019 Real-Time Live!, pp. 2. Cited by: §1.
  • [32] T. Park, M. Liu, T. Wang, and J. Zhu (2019) Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2337–2346. Cited by: §1.
  • [33] S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, and H. Lee (2016) Generative adversarial text to image synthesis. arXiv preprint arXiv:1605.05396. Cited by: §1, §2, §2.
  • [34] A. Shrivastava and A. Gupta (2016) Contextual priming and feedback for faster r-cnn. In European Conference on Computer Vision, pp. 330–348. Cited by: §1.
  • [35] V. Steinbiss, B. Tran, and H. Ney (1994) Improvements in beam search. In Third International Conference on Spoken Language Processing, Cited by: §3.3.
  • [36] A. Torralba and P. Sinha (2001) Statistical context priming for object detection. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, Vol. 1, pp. 763–770. Cited by: §1.
  • [37] A. van den Oord and N. Kalchbrenner (2016) Pixel rnn. Cited by: §2.
  • [38] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin (2017) Attention is all you need. In Advances in neural information processing systems, pp. 5998–6008. Cited by: §3.2.
  • [39] K. Wang, Y. Lin, B. Weissmann, M. Savva, A. X. Chang, and D. Ritchie (2019) Planit: planning and instantiating indoor scenes with relation graph and spatial prior networks. ACM Transactions on Graphics (TOG) 38 (4), pp. 1–15. Cited by: §2.
  • [40] K. Wang, M. Savva, A. X. Chang, and D. Ritchie (2018) Deep convolutional priors for indoor scene synthesis. ACM Transactions on Graphics (TOG) 37 (4), pp. 70. Cited by: §1, §2.
  • [41] T. Wang, M. Liu, J. Zhu, A. Tao, J. Kautz, and B. Catanzaro (2018) High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Cited by: §2.
  • [42] J. Wu, E. Lu, P. Kohli, W. T. Freeman, and J. B. Tenenbaum (2017) Learning to see physics via visual de-animation. In Advances in Neural Information Processing Systems, Cited by: §1.
  • [43] J. Wu, J. B. Tenenbaum, and P. Kohli (2017) Neural scene de-rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §1.
  • [44] X. Yang, M. E. Yümer, P. Asente, M. Kraley, D. Kifer, and C. L. Giles (2017)

    Learning to extract semantic structure from documents using multimodal fully convolutional neural network

    CoRR abs/1706.02337. External Links: Link, 1706.02337 Cited by: §1.
  • [45] A. J. Yepes, J. Tang, and X. Zhong PubLayNet: largest dataset ever for document layout analysis. Cited by: 4th item, §2, §4.1, §4.1.
  • [46] H. Zhang, I. Goodfellow, D. Metaxas, and A. Odena (2018) Self-attention generative adversarial networks. arXiv preprint arXiv:1805.08318. Cited by: §1.
  • [47] H. Zhang, T. Xu, H. Li, S. Zhang, X. Wang, X. Huang, and D. N. Metaxas (2017) Stackgan: text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision, pp. 5907–5915. Cited by: §1, §2, §2.
  • [48] B. Zhao, L. Meng, W. Yin, and L. Sigal (2019) Image generation from layout. In CVPR, Cited by: Figure 1, §1, §1, §2, §4.5, Table 3, Table 3.
  • [49] X. Zheng, X. Qiao, Y. Cao, and R. W. Lau (2019) Content-aware generative modeling of graphic design layouts. ACM Transactions on Graphics (TOG) 38 (4), pp. 1–15. Cited by: §2, §2.

Appendix 0.A Architecture and training details

In all our experiments, our base model consists of , , , and where is the number of grid cells (corresponding to locations and size of bounding boxes), is the dimension of embedding in each of the layers, is the number of self-attention layers, is the number of heads in each of the self-attention layer, and is the number of units in feedforward layer that follows the self-attention part in a single decoder block. We also use a dropout of at the end of each feedforward layer for regularization. Starting with start token , our model predicts category, location and shape of next bounding box in a raster scan order. We fix the the maximum number of elements in each of the datasets to which covers over 99.9% of the layouts in each of the COCO, Rico and PubLayNet datasets.

We also used Adam optimizer [21]

with Noam learning rate scheduling. We train our model for 20 epochs for each dataset with early stopping based on maximum log likelihood on validation layouts (overall we trained our model for 8 epochs on COCO, 12 epochs on Rico, and 16 epochs on PubLayNet). Our COCO Bounding Boxes model takes about 2 hours to train on a single NVIDIA GTX1080 GPU. Batching matters a lot to improve the training speed. We want to have evenly divided batches, with absolutely minimal padding. We sort the layouts by the number of elements and search over this sorted list to use find tight batches for training.

(a) Ours
(b) LayoutVAE [18]
(c) SG2IM [17]
(d) Obj-GAN [26]
Figure 9: Layout samples of COCO bounding boxes. LayoutVAE is constrained to start with a fixed set of categories. For the set of categories, which rarely occur together, LayoutVAE might have hard time finding appropriate placement for them. For example, in second layout of bottom row (of LayoutVAE), clock and traffic lights are put together which are a bit unlikely. Our method doesn’t face this problem. One artifact of our method is since it generates most likely layouts first (using greedy decoding), it might not end up using some less common objects (although one work around of this problem by using penalized beam search or nucleus sampling [14]). Another point to be noted is both LayoutVAE and SG2IM needs as input a set of categories to be placed in the layout. For above samples, we use validation data to get the input sets.
(a) Ours
(b) LayoutVAE [18]
Figure 10: Layout samples of Rico. LayoutVAE’s bounding boxes are not aligned with each other as is the case with real samples (and our samples)
(a) Ours
(b) LayoutVAE [18]
(c) SG2IM [17]
(d) Obj-GAN [26]
Figure 11: Layout samples of PubLayNet. ObjGAN, SG2IM and LayoutVAE’s bounding boxes are not aligned with each other as is the case with real samples (and our samples)

Appendix 0.B Evaluation Details

In this section, we provide more details on various qualitative and qualitative evaluation done in Section 4. We also provide some additional results (that didn’t fit in the paper).

0.b.1 Generated samples

We show random samples generated for each of the dataset using different methods listed in Section 4.2. In case of LayoutVAE [18], we use label set of layouts in validation dataset as input to generate samples. In case of LayoutTransformer, we take one layout element from validation dataset and complete the layout using our model.

0.b.2 Computing nearest neighbors

While there is no standard method for comparing two layouts, in Section 4.3 of the paper, we use modified Chamfer distance111 in order to compare two layouts (with different number of elements). This metric doesn’t take into account categories of various layout elements and just compute euclidean distance between bounding box top left coordinates and height and width.

0.b.3 Computing log-likelihood

In order to compute NLL for layouts generated by LayoutVAE, we follow the approach as provided by the authors in their paper. Using teacher forcing, for a layout in validation set we compute Monte Carlo estimate of NLL by drawing 1000 samples from conditional prior. We add the NLL for CountVAE and BBoxVAE.

Figure 12: We observe the impact of operations such as left right flip, and up down flip on log likelihood of the layout. We observe that unlikely layouts (such as fog at the bottom of image have higher NLL than the layouts from data.

0.b.4 Layout Verification

Since in our method it is straightforward to compute likelihood of a layout, we can use our method to test if a given layout is likely or not. Figure 12 shows the NLL given by our model by doing left-right and top-down inversion of layouts in COCO (following  [25]). In case of COCO, if we flip a layout left-right, we observe that layout remains likely, however flipping the layout upside decreases the likelihood (or increases the NLL of the layout). This is intuitive since it is unlikely to see fog in the bottom of an image, while skis on top of a person.

0.b.5 LayoutGAN

LayoutGAN [25]

represents each layout with a fixed number of bounding boxes. Starting with bounding box coordinates sampled from a Gaussian distribution, its GAN based framework assigns new coordinates to each bounding box to resemble the layouts from given data. Optionally, it uses non-maximum suppression (NMS) to remove duplicates. The problem setup in LayoutGAN is similar to the proposed approach and they do not condition the generated layout on anything. Like many GAN setups, LayoutGAN is non-trivial to train across multiple datasets. In our implementation of LayoutGAN, we were unable to prevent mode collapse for all datasets except for MNIST.

(a) LayoutGAN – COCO
(b) LayoutGAN – Rico
(c) LayoutGAN – PubLayNet
Figure 13: Layout samples using LayoutGAN. We were unable to prevent mode collapse for all datasets except for MNIST.

Appendix 0.C Dataset Statistics

In this section, we share statistics of different elements and their categories in our dataset. In particular, we share the total number of occurrences of an element in the training dataset (in descending order) and the total number of distinct layouts an element was present in throughout the training data. Table 12 show these statistics for COCO bounding boxes, Tables 1010 show the statistics for Rico wireframes, and table 11 show the statistics for PubLayNet documents.

Category # occurrences # layouts
Text 387457 50322
Image 179956 38958
Icon 160817 43380
Text Button 118480 33908
List Item 72255 9620
Input 18514 8532
Card 12873 3775
Web View 10782 5808
Radio Button 4890 1258
Drawer 4138 4136
Checkbox 3734 1126
Advertisement 3695 3365
Category # occurrences # layouts
Modal 3248 3248
Pager Indicator 2041 1528
Slider 1619 954
On/Off Switch 1260 683
Button Bar 577 577
Toolbar 444 395
Number Stepper 369 147
Multi-Tab 284 275
Date Picker 230 217
Map View 186 94
Video 168 144
Bottom Navigation 75 27
Table 10: Category statistics for Rico
Category # occurrences # layouts
text 2343356 334548
title 627125 255731
figure 109292 91968
table 102514 86460
list 80759 53049
Table 11: Category statistics for PubLayNet
Category # occurrences # layouts
person 257253 64115
other 117266 117266
car 43533 12251
chair 38073 12774
tree 36466 36466
sky-other 31808 31808
wall-concrete 31481 31481
clothes 27657 27657
book 24077 5332
bottle 24070 8501
building-other 23021 23021
grass 22575 22575
metal 22526 22526
cup 20574 9189
wall-other 19095 19095
pavement 18311 18311
furniture-other 17882 17882
table 16282 16282
dining table 15695 11837
road 15402 15402
bowl 14323 7111
window-other 14209 14209
textile-other 13052 13052
traffic light 12842 4139
handbag 12342 6841
light 11772 11772
fence 11303 11303
umbrella 11265 3968
plastic 11137 11137
boat 10576 3025
ceiling-other 10546 10546
bird 10542 3237
dirt 10163 10163
truck 9970 6127
clouds 9886 9886
bush 9849 9849
bench 9820 5570
plant-other 9522 9522
paper 9521 9521
door-stuff 9475 9475
sheep 9223 1529
banana 9195 2243
floor-other 8893 8893
kite 8802 2261
backpack 8714 5528
motorcycle 8654 3502
potted plant 8631 4452
cow 8014 1968
wine glass 7839 2533
knife 7760 4326
carrot 7758 1683
broccoli 7261 1939
cabinet 7176 7176
bicycle 7056 3252
donut 7005 1523
food-other 6672 6672
wall-wood 6642 6642
skis 6623 3082
Category # occurrences # layouts
floor-tile 6618 6618
sea 6598 6598
vase 6577 3593
horse 6567 2941
house 6549 6549
tie 6448 3810
cell phone 6422 4803
floor-wood 6324 6324
clock 6320 4659
orange 6302 1699
sports ball 6299 4262
cake 6296 2925
ground-other 6252 6252
spoon 6159 3529
suitcase 6112 2402
surfboard 6095 3486
bus 6061 3952
pizza 5807 3166
tv 5803 4561
couch 5779 4423
apple 5776 1586
remote 5700 3076
sink 5609 4678
skateboard 5536 3476
dog 5500 4385
elephant 5484 2143
fork 5474 3555
wall-tile 5290 5290
zebra 5269 1916
playingfield 5251 5251
wall-brick 5246 5246
airplane 5129 2986
giraffe 5128 2546
snow 5114 5114
curtain 5101 5101
wood 5053 5053
laptop 4960 3524
mountain 4887 4887
carpet 4858 4858
tennis racket 4807 3394
cat 4766 4114
teddy bear 4729 2140
sand 4688 4688
counter 4589 4589
shelf 4589 4589
train 4570 3588
roof 4490 4490
sandwich 4356 2365
bed 4192 3682
toilet 4149 3353
banner 4135 4135
cardboard 3787 3787
baseball glove 3747 2629
mirror-stuff 3622 3622
rock 3397 3397
oven 3334 2877
baseball bat 3273 2506
Category # occurrences # layouts
flower 3259 3259
leaves 3169 3169
cloth 3129 3129
structural-other 3016 3016
cage 2911 2911
desk-stuff 2909 2909
hot dog 2884 1222
keyboard 2854 2115
branch 2813 2813
railroad 2720 2720
rug 2703 2703
frisbee 2681 2184
snowboard 2681 1654
stairs 2667 2667
fog 2659 2659
refrigerator 2634 2360
gravel 2613 2613
blanket 2598 2598
towel 2558 2558
hill 2498 2498
water-other 2453 2453
wall-panel 2357 2357
river 2313 2313
window-blind 2297 2297
mouse 2261 1876
fruit 2112 2112
railing 2068 2068
wall-stone 2020 2020
vegetable 2016 2016
platform 2009 2009
skyscraper 1998 1998
pillow 1986 1986
stop sign 1983 1734
toothbrush 1945 1007
fire hydrant 1865 1711
stone 1828 1828
bridge 1676 1676
microwave 1672 1547
tent 1486 1486
scissors 1464 947
napkin 1405 1405
straw 1385 1385
net 1362 1362
bear 1294 960
parking meter 1283 705
floor-stone 1259 1259
cupboard 1004 1004
floor-marble 1002 1002
solid-other 749 749
mud 659 659
mat 559 559
salad 477 477
ceiling-tile 351 351
moss 256 256
toaster 225 217
hair drier 198 189
waterdrops 121 121
Table 12: Category statistics for COCO