visual-compatibility
Context-Aware Visual Compatibility Prediction (https://arxiv.org/abs/1902.03646)
view repo
How do we determine whether two or more clothing items are compatible or visually appealing? Part of the answer lies in understanding of visual aesthetics, and is biased by personal preferences shaped by social attitudes, time, and place. In this work we propose a method that predicts compatibility between two items based on their visual features, as well as their context. We define context as the products that are known to be compatible with each of these item. Our model is in contrast to other metric learning approaches that rely on pairwise comparisons between item features alone. We address the compatibility prediction problem using a graph neural network that learns to generate product embeddings conditioned on their context. We present results for two prediction tasks (fill in the blank and outfit compatibility) tested on two fashion datasets Polyvore and Fashion-Gen, and on a subset of the Amazon dataset; we achieve state of the art results when using context information and show how test performance improves as more context is used.
READ FULL TEXT VIEW PDFContext-Aware Visual Compatibility Prediction (https://arxiv.org/abs/1902.03646)
Fashion-AI is a PyTorch code base that implements various sate-of-the-art algorithms related to fashion AI, e.g., compatibility prediction, outfit recommendation, etc.
Predicting fashion compatibility refers to the task of determining whether a set of fashion items go well together. In its ideal form, it involves understanding the visual styles of garments, being cognizant of social and cultural attitudes, and making sure that when worn together the outfit is aesthetically pleasing. The task is fundamental to a variety of industry applications such as personalized fashion design [19], outfit composition [7], wardrobe creation [16], item recommendation [31] and fashion trend forecasting [1]. Fashion compatibility, however, is a complex task that depends on subjective notions of style, context, and trend – all properties that may vary from one individual to another and evolve over time.
Previous work [24, 36] on the problem of fashion compatibility prediction uses models that mainly perform pairwise comparisons between items based on item information such as image, category, description, …, etc. These approaches have the drawback that each pair of items considered are treated independently, making the final prediction rely on comparisons between the features of each item in isolation. In such a comparison mechanism that discards context, the model makes the same prediction for a given pair of clothing items every time. For example, if the model is trained to match a specific style of shirt with a specific style of shoes, it will consistently make this same prediction every time. However, as compatibility is a subjective measure that can change with trends and across individuals, such inflexible behaviour is not always desirable at test time. The compatibility between the aforementioned shirt and shoes is not only defined by the features of these items alone, but is also biased by the individual’s preferences and sense of fashion. We thus define the context of a clothing item to be the set of items that it is compatible with, and address the limitation of inflexible predictions by introducing a model that makes compatibility decisions based on the visual features, as well as the context of each item. This consideration gives the model some background as to what we consider “compatible”, in itself a subjective bias of the individual and the trend of the time.
In this paper, we propose to leverage the underlying relational information between items in a collection to make better compatibility predictions. We use fashion as our theme, and represent clothing items and their pairwise compatibility as a graph, where vertices are the fashion items and edges connect pairs of items that are compatible; we then use a graph neural network based model to learn to predict edges. Our model is based on the graph auto-encoder framework [22], which defines an encoder that computes node embeddings and a decoder that is applied on the embedding of each product. Graph auto-encoders have previously been used for related problems such as recommender systems [35], and we extend the idea to the fashion compatibility prediction task. The encoder part of the model computes item embeddings depending on their connections, while the decoder uses these embeddings to compute the compatibility between item pairs. By conditioning the embeddings of the products on the neighbours, the style information contained in the representation is more robust, and hence produces more accurate compatibility predictions. This accuracy is tested by a set of experiments we perform on three datasets: Polyvore [12], Fashion-Gen [28] and Amazon [24], and through two tasks (1) outfit completion (see Section 4.1) and (2) outfit compatibility prediction (see Section 4.1). We compare our model with previous methods and obtain state of the art results. During test time, we provide our model with varying amount of context of each item being tested and empirically show, in addition, that the more context we use, the more accurate our predictions get.
This work has the following main contributions, (1) we propose the first fashion compatibility method that uses context information; (2) we perform an empirical study of how the amount of neighbourhood information used during test time influences the prediction accuracy; and (3) we show that our method outperforms other baseline approaches that do not use the context around each item on the Polvvore [12], Fashion-Gen [28], and Amazon [24] datasets.
As our proposed model uses graph neural networks to perform fashion compatibility prediction, we group previous work related to our proposed model into two categories that we discuss in this section. In what follows, an outfit is a set of clothing items that can be worn concurrently. We say that an outfit is compatible, if the clothing items composing the outfit are aesthetically pleasing when worn together; we extend an outfit when we add clothing item(s) to the set composing the outfit.
To approach the task of visual compatibility prediction, McAuley et al. [24] learn a compatibility metric on top of CNN-extracted visual features, and apply their method to pairs of products such that the learned distance in the embedding space is interpreted as compatibility. Their approach is improved by Veit et al. [38], who instead of using pre-computed features for the images, use an end-to-end siamese network to predict compatibility between pairs of images. A similar end-to-end approach [19]
shows that jointly learning the feature extractor and the recommender system leads to better results. The evolution of fashion style has an important role in compatibility estimation, and He
et al. [14] study how previous methods can be adapted to model the visual evolution of fashion trends within recommender systems.Some variations of this task include predicting the compatibility of an outfit, to generate outfits from a personal closet [34] for example, or determining the item that best extends a partial outfit. To approach these tasks, Han et al. [12] consider a fashion outfit to be an ordered sequence of products and use a bidirectional LSTM on top of the CNN-extracted features from the images and semantic information extracted from text in the embedding space. This method was improved by adding a new style embedding for the full outfit [27]. Vasileba et al. [36] also use textual information to improve the product embeddings, along with using conditional similarity networks [37] to produce type-conditioned embeddings and learn a metric for compatibility. This approach projects each product embedding to a new space, depending on the type of the item pairs being compared.
Extending neural networks to work with graph structured data was first proposed by Gori et al. [10] and Scarselli et al. [29]. The interest in this topic resurged recently, with the proposal of spectral graph neural networks [5] and its improvements [6, 21]. Gilmer et al. [9] showed that most of the methods that apply neural networks to graphs [25, 39, 11] can be seen as specific instances of a learnable message passing framework on graphs. For an in-depth review of different approaches that apply neural networks to graph-structured data, we refer the reader to the work by Bronstein et al. [4] and Battaglia et al. [2]
, which explores how relational inductive biases can be injected in deep learning architectures.
Graph neural networks have been applied to product recommendation, which is similar to product compatibility prediction. In this task, the goal is to predict compatibility between users and products (as opposed to a pair of products). Van den Berg et al. [35] showed how this task can be approached as a link prediction problem in a graph. Similarly, graphs can also be used to take advantage of the structure within the rows and columns of a matrix completion problem applied to product recommendation [18, 26]. Recently, a graph-based recommender system has been scaled to web-scale [40], operating on a graph with more than 3 billion nodes consisting of pins and boards from Pinterest.
The approach we use in this work is similar to the metric learning idea of Vasileba et al. [36], but rather than using text to improve products embeddings, we use a graph to exploit structural information and obtain better product embeddings. Our model is based on the graph auto-encoder (GAE) framework defined by Kipf et al. [22], which has been used for tasks like knowledge base completion [30] and collaborative filtering [35]. In this framework, the encoder gets as input an incomplete graph, and produces an embedding for each node. Then, the node embeddings are used by the decoder to predict the missing edges in the graph.
Let be an undirected graph with nodes and edges
connecting pairs of nodes. Each node in the graph is represented with a vector of features
, and is a matrix that contains the features of all nodes in the graph. Each row of , denoted as , contains the features of one node, i.e. represent the features of the node. The graph is represented by an adjacency matrix , where if there exist an edge between nodes and and otherwise.The objective of the model is to learn an encoding and a decoding function. The encoder transforms the initial features into a new representation , depending on the structure defined by the adjacency matrix . This new matrix follows the same structure as the initial matrix , so the -th row contains the new features for the -th node. Then, the decoder uses the new representations to reconstruct the adjacency matrix. This whole process can be seen as encoding
the input features to a new space, where the distance between two points can be mapped to the probability of whether or not an edge exists between them. We use a
decoder to compute this probability using the features of each node: , which for our purposes represents the compatibility between items and .In this work, the encoder is a Graph Convolutional Network (Section 3.1) and the decoder (Section 3.2) learns a metric to predict the compatibility score between pairs of products . Figure 2 shows a scheme of how this encoder-decoder mechanism works.
From the point of view of a single node , the encoder will transform its initial visual features into a new representation . The initial features, which can be computed with a CNN as a feature extractor, contain information about how an item looks like, e.g., shape, color, size. However, we want the new representation produced by the encoder to capture not only the product properties but also structural information about the other products it is compatile with. In other words, we want the new representation of each node to contain information about itself, but also about its neighbours , where denotes the set of nodes that are connected to node . Therefore, the encoder is a function that aggregates the local neighbourhood around a node to include neighbourhood information in the learned representations. This function is implemented as a deep Graph Convolutional Network (GCN) [21] that can have several hidden layers. Thus, the final value of is a composition of the functions computed at each hidden layer, which produces hidden activations . A single layer takes the following form.
(1) |
Here, is the input of the -th node at layer , and is its output. In its matrix form, the function operates on all the nodes of the graph at the same time:
(2) |
Here, for the first layer. We denote as the normalized -th step adjacency matrix, where contains self-connections, and contains first step neighbours with self-connections. We let , normalizing it row-wise using the diagonal degree matrix . Context information is controlled by the parameter that represents the depth of the neighbourhood that is being considered during training: the neighbourhood at depth of node is the set of all nodes that are at distance (number of edges traveled) at most from . We let for all our experiments, meaning that we only use neighbours at depth one in each layer. is a matrix, which contains the trainable parameters for layer
. We apply techniques such as batch normalization
[17], dropout [33] or weight regularization at each layer.Finally, we introduce a regularization technique applied to the matrix , which consists of randomly removing all the incident edges of some nodes with a probability . The goal of this technique is two-fold: (1) it introduces some changes in the structure of the graph, making it more robust against changes in structure, and (2) it trains the model to perform well for nodes that do not have neighbours, making it more robust to scenarios with low relational information.
We want the decoder to be a function that computes the probability that two nodes are connected. This scenario is known as metric learning [3], where the goal is to learn a notion of similarity or compatibility between data samples. It is relevant to note that similarity and compatibility are not exactly the same. Similarity measures how similar two nodes are, for example two shirts might be similar because they have the same shape and color, but they are not necessarily compatible. Compatibility is a property that measures how well two items go together.
In its general form, metric learning can be defined as learning a function that represents the distance between two -dimensional vectors. Therefore, our decoder function takes inspiration from other metric learning approaches [23, 15, 32]. In our case, we want to train the decoder to model the compatibility between pairs of items, so we want the output of to be bounded by the interval .
The decoder function we use is similar to the one proposed by [8]. Given the representations of two nodes and computed with the encoder model described above, the decoder outputs the probability that these two nodes are connected by an edge.
(3) |
Here is absolute value, and and are learnable parameters.
is the sigmoid function that maps a scalar value to a valid probability
.The form of the decoder described in Equation 3
can be seen as a logistic regression decoder operating on the absolute difference between the two input vectors. The absolute value is used to ensure that the decoder is symmetric,
i.e., the output of and is the same, making it invariant to the order of the nodes.
![]() |
![]() |
![]() |
![]() |
The model is trained to predict compatibility among the products. With being the adjacency matrix of the graph of items, we randomly remove a subset of edges to generate an incomplete adjacency matrix . The set of edges removed is denoted by , as they represent positive edges, i.e., pairs of nodes such that . We then randomly sample a set of negative edges , which represent pairs of nodes that are not connected, i.e., products that are not compatible. The model is trained to predict the edges that contain both positive and negative edges. Therefore, given the incomplete adjacency matrix and the initial features for each node , the decoder predicts the edges defined in , and the model is optimized by minimizing the cross entropy loss between the predicted edges and their ground truth values, which is for the edges in and for the edges in .
We apply our model to two tasks that can be recast as a graph edge prediction problem. In what follows, we let denote the set of fashion items in a given outfit, and denote the edge between nodes and .
The fill-in-the-blank task consists of choosing the item that best extends an outfit from among a given set of possible item choices. We follow the setup described in Han et al. [12], where one FITB question is defined for each test outfit. Each question consists of a set of products that form a partial outfit, and a set of possible choices that includes the correct answer and randomly chosen products. In our experiments we set the number of choices to 4. An example of one of these questions can be seen in Figure 2(a), where the top row shows the products of a partial outfit and the bottom row shows the possible choices for extending it. FITB can be framed as an edge prediction problem where the model first generates the probability of edges between item pairs for all and . Then, the score for each of the choices is computed as , and the one with the highest score is the item that is selected to be added to the partial outfit. The task itself is evaluated using the same metric defined by Han et al. [12]: by measuring whether or not the correct item was selected from the list of choices.
In the outfit compatibility prediction task, the goal is to produce an outfit compatibility score, which represents the overall compatibility of the items forming the outfit. Scores close to represent compatible outfits, and scores close to represent incompatible outfits. The task can be framed as an edge prediction problem where the model predicts the probability of every edge between all possible item pairs; this means predicting the probability of edges for each outfit. The compatibility score of the outfit is the average over all pairwise edge probabilities . The outfit compatibility prediction task is evaluated using the area under the ROC curve for the predicted scores.
Let the -neighbourhood of node in our relational graph be the set of nodes that are visited by a breadth-first-search process, starting from . In order to measure the effect of the size of relational structure around each item, during testing we let each test sample contain the items and their -neighbourhoods, and we evaluate our model by varying . Thus, when (Figure 3(a)) no relational information is used, and the embedding of each product is based only on its own features. As the value of increases (Figures 3(b) and 3(c)), the embedding of the items compared will be conditioned on more neighbours. Note that this is applied only at evaluation time; during training, we use all available edges. For all results in the following sections we report the value of used for each experiment.
We test our model on three datasets, as well as on a few of their variations that we discuss below.
The Polyvore dataset [12] is a crowd-sourced dataset created by the users of a website of the same name; the website allowed its members to upload photos of fashion items, and collect them into outfits. It contains a total of items that form different outfits. The maximum number of items per outfit is , and the average number of items per outfit is . The graph is created by connecting each pair of nodes that appear in the same outfit with an edge. We train our model with the train set of the Polyvore dataset, and test it on a few variations obtained from this dataset, described below.
The FITB task contains questions and the outfit compatibility task has valid, and invalid outfits. In the original Polyvore dataset, the wrong FITB choices and the invalid outfits are selected randomly from among all remaining products. The resampled dataset proposed by Vasileba et al. [36] is more challenging: the incorrect choices in each question of the FITB task are sampled from the items having the same category as the correct choice; for outfit compatibility, outfits are sampled randomly such that each item in a given outfit is from a distinct category. We also propose a more challenging set which we call subset where we limit the outfits size to 3 randomly selected items. In this scenario the tasks become harder because less information is available to the model.
Fashion-Gen [28] is a dataset of fashion products collected from an online platform that sells luxury goods from independent designers. Each product has images, descriptions, attributes, and relational information. Fashion-Gen relations are defined by professional designers and adhere to a general theme, while Polyvore’s relations are generated by users with different tastes and notions of compatibility.
We created outfits from Fahion-Gen by grouping between and products that are connected together. The training set consists of different outfits from the collections , and the validation and test sets have and outfits respectively, from the collection. The incorrect FITB choices and the invalid outfits for the compatibility task are randomly sampled items that satisfy gender and category restrictions, as in the case of the resampled Polyvore dataset.
The Amazon products dataset [24, 14] contains over million relationships between almost million products of different categories. In this work we focus on the clothing products, and we apply our method to the Men and Women categories. There are types of relationships between items: (1) users who viewed also viewed ; (2) users who viewed bought ; (3) users who bought also bought ; and (4) users bought and simultaneously. For the latter two cases, we make the assumption that the pair or items and are compatible and evaluate our model based on this assumption. We evaluate our model by predicting the latter two, since they indicate products that might be complementary [24]. We use the features they provide, which are computed with a CNN.
![]() |
![]() |
![]() |
Our model has graph convolutional layers with , units, dropout of applied at the input and batch normalization at its output. The value of applied to is . The input to each node are -dimensional feature vectors extracted with a ResNet-50 [13]
from the image of each product, and are normalized to zero-mean and unit variance. It is trained with Adam
[20], with a learning rate of for iterations with early stopping.The Siamese Network
baseline is trained with triplets of compatible and incompatible pairs of items. It consists on a ImageNet pretrained ResNet-50 at each branch and a metric learning output layer. We train it using SGD with a learning rate of
and a momentum of .
![]() |
![]() |
![]() |
![]() |
We report our results for this task in Table 1. The first three rows correspond to previous work, and the following three rows show the scores obtained by our model for different values of . As shown in the table, the scores consistently increases with , from of accuracy with to with . This behaviour is better seen in Figure 4(a) which shows how the accuracy in the FITB task increases as a function of . When the other methods perform better, because without structure our model is simpler. However, we can see how as more neighbourhood information is used, the results in the FITB task increase, which shows that using information from neighbouring nodes is a useful approach if extra relational information is available.
FITB Accuracy | Compat. AUC | |||
Method | Orig. | Res. | Orig. | Res. |
Siamese Net [36] | 54.2 | 54.4 | 0.85 | 0.85 |
Bi-LSTM [12] | 68.6 | 64.9 | 0.90 | 0.94 |
TA-CSN [36] | 86.1 | 65.0 | 0.98 | 0.93 |
Ours (k = 0) | 62.2 | 47.0 | 0.86 | 0.76 |
Ours (k = 3) | 95.9 | 90.9 | 0.99 | 0.98 |
Ours (k = 15) | 96.9 | 92.7 | 0.99 | 0.99 |
Ours (k = 0)† | 59.5 | 45.3 | 0.69 | 0.64 |
Ours (k = 3)† | 79.1 | 69.4 | 0.92 | 0.90 |
Ours (k = 15)† | 88.2 | 82.1 | 0.93 | 0.92 |
For the resampled setup, the accuracy also increases with , going from to , which is lower than its original counterpart, showing that the resampled task is indeed more difficult.
The last rows of Table 1 (marked with †) correspond to this scenario, and we can see that compared to when using the full outfit, the the FITB accuracy drops from to for the original version, and from to for the resampled version, both at .
Method | FITB Acc. | Compatibility AUC |
Siamese Network | 56.3 | 0.69 |
Ours (k = 0) | 51.9 | 0.72 |
Ours (k = 3) | 65.0 | 0.84 |
Ours (k = 15) | 76.1 | 0.90 |
Ours (k = 30) | 77.1 | 0.91 |
The results for the FITB task on the Fashion-Gen dataset are shown in Table 2 as a function of . Similar to the results for variations of Polyvore, we see in Figure 4(c) how an increase in the value of improves the performance of our model also for the Fashion-Gen dataset. For example, it increases by points by using up to neighbourhood nodes for each item, compared to using no neighbourhood information at all. When compared to the Siamese Network baseline, we observe how the siamese model is better than our model without structure, but with our method outperforms the baseline.
Table 1 shows the results obtained by our model on the compatibility prediction task for different values of . Similarly to the previous task, results show that using more neighborhood information improves the performance on the outfit compatibility task, where the AUC increases from with to with .
The scores on the resampled version are similar to the original version, increasing the AUC from to with a larger value for .
The results on this test data is denoted with † in the table, and we see how in this scenario the scores decrease from to and for the original and resampled tasks respectively, both with . As with the FITB task, here we observe again how using extra information in the form of relations with other products is beneficial to achieve better performance.
The results on this task for the Fashion-Gen outfits dataset are shown in the second column of Table 2, for different values of . As can be seen, the larger the value of , the better the performance. This trend is better shown in Figure 4(d), where we can see how increasing from to steadily improves the performance, and plateaus afterwards.
![]() |
![]() |
With the above experiments, we have seen how increasing the amount of neighbourhood information improves the results on all tasks. To better understand the role of context, we use an example from Polyvore to demonstrate how the context of an item can influence its predicted compatibility with another product. Figure 6 shows the compatibility predicted between a pair of trousers and two pairs of shoes depending on two different contexts. Figure 5(a) shows the original context of the trousers, and the shoes selected are the correct ones. However, if we change the context of the trousers to a different set of clothes, as in Figure 5(b), the outcome of the prediction is now a different pair of shoes (more formal one) that presumably are a better match given the new context.
We also evaluate how our method can be applied to predict relations between products in the Amazon dataset. We train a model for each type of relationship and also evaluate how one model trained with clothes from one gender transfers to the other gender. This cross-gender setup allows us to evaluate how the model adapts to changes in context, as opposed to a baseline that ignores context altogether. In Table 3 we show that our model achieves state of the art results for the ’also bought’ relation, and similar results for the ’bought together’ relation. The ’bought together’ relationship has much less connections than the ’also bought’, so our model is less effective at using context to improve the results. However, since in that scenario the model has been trained with less connections, it performs better with , because it is more similar to the training behaviour. In Table 4 we show the results of one model trained with men’s clothing and tested with women’s clothing (and vice versa). The model denoted with † does not use relational information during training and testing, so is the baseline for not using contextual information (). As it can be seen, the more neighbourhood information a model uses, the most robust it is to the domain change. This occurs because when the model relies on context, it can adapt better to unseen styles or clothing types.
Also bought | Bought together | |||
Method | Men | Women | Men | Women |
McAuley et al. [24] | 93.3 | 91.2 | 95.1 | 94.3 |
Ours (k = 0) | 57.9 | 53.8 | 79.5 | 71.7 |
Ours (k = 3) | 92.6 | 92.9 | 94.5 | 94.5 |
Ours (k = 10) | 97.1 | 95.8 | 94.0 | 94.8 |
|
|
In this paper we have seen how context information can be used to improve the performance on compatibility prediction tasks using a graph neural network based model. We experimentally show that increasing the amount of context improves the performance of our model on all tasks. We conduct experiments on three different fashion datasets and obtain state of the art results when context is used during test time.
AAAI Conference on Artificial Intelligence (AAAI)
, 2018.Journal of Machine Learning Research (JMLR)
, 15(1):1929–1958, 2014.
Comments
There are no comments yet.