Pre-training without Natural Images

01/21/2021 ∙ by Hirokatsu Kataoka, et al. ∙ 23

Is it possible to use convolutional neural networks pre-trained without any natural images to assist natural image understanding? The paper proposes a novel concept, Formula-driven Supervised Learning. We automatically generate image patterns and their category labels by assigning fractals, which are based on a natural law existing in the background knowledge of the real world. Theoretically, the use of automatically generated images instead of natural images in the pre-training phase allows us to generate an infinite scale dataset of labeled images. Although the models pre-trained with the proposed Fractal DataBase (FractalDB), a database without natural images, does not necessarily outperform models pre-trained with human annotated datasets at all settings, we are able to partially surpass the accuracy of ImageNet/Places pre-trained models. The image representation with the proposed FractalDB captures a unique feature in the visualization of convolutional layers and attentions.



There are no comments yet.


page 2

page 4

page 5

page 9

Code Repositories


Pre-training without Natural Images (ACCV 2020 Best Paper Honorable Mention Award)

view repo
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

The introduction of sophisticated pre-training image representation has lead to a great expansion of the potential of image recognition. Image representations with e.g., the ImageNet/Places pre-trained convolutional neural networks (CNN), has without doubt become the most important breakthrough in recent years Deng et al. (2009); Zhou et al. (2017). We had lots to learn from the ImageNet project, such as huge amount of annotations done by crowdsourcing and well-organized categorization based on WordNet Fellbaum (1998). However, due to the fact that the annotation was done by a large number of unspecified people, most of whom are unknowledgeable and not experts in image classification and the corresponding areas, the dataset contains mistaken, privacy-violated, and ethics-related labels Buolamwini and Gebru (2018); Yang et al. . This limits the ImageNet to only non-commercial usage because the images included in the dataset does not clear the right related issues. We believe that this aspect of pre-trained models significantly narrows down the prospects of vision-based recognition.

We begin by considering what a pre-trained CNN model with a million natural images is. In most cases, representative image datasets consist of natural images taken by a camera that express a projection of the real world. Although the space of image representation is enormous, a CNN model has been shown to be capable of recognition of natural images from among around one million natural images from the ImageNet dataset. We believe that labeled images on the order of millions have a great potential to improve image representation as a pre-trained model. However, at the moment, a curious question occurs:

Can we accomplish pre-training without any natural images for parameter fine-tuning on a dataset including natural images? To the best of our knowledge, the ImageNet/Places pre-trained models have not been replaced by a model trained without natural images. Here, we deeply consider pre-training without natural images. In order to replace the models pre-trained with natural images, we attempt to find a method for automatically generating images. Automatically generating a large-scale labeled image dataset is challenging, however, a model pre-trained without natural images makes it possible to solve problems related to privacy, copyright, and ethics, as well as issues related to the cost of image collection and labeling.

Unlike a synthetic image dataset, could we automatically make image patterns and their labels with image projection from a mathematical formula? Regarding synthetic datasets, the SURREAL dataset Varol et al. (2017)

has successfully made training samples of estimating human poses with human-based motion capture (mocap) and background. In contrast, our Formula-driven Supervised Learning and the generated formula-driven image dataset has a great potential to automatically generate an image pattern and a label. For example, we consider using

fractals, a sophisticated natural formula Mandelbrot (1983). Generated fractals can differ drastically with a slight change in the parameters, and can often be distinguished in the real-world. Most natural objects appear to be composed of complex patterns, but fractals allow us to understand and reproduce these patterns.

We believe that the concept of pre-training without natural images can simplify large-scale DB construction with formula-driven image projection in order to efficiently use a pre-trained model. Therefore, the formula-driven image dataset that includes automatically generated image patterns and labels helps to efficiently solve some of the current issues involved in using a CNN, namely, large-scale image database construction without human annotation and image downloading. Basically, the dataset construction does not rely on any natural images (e.g. ImageNet Deng et al. (2009) or Places Zhou et al. (2017)) or closely resembling synthetic images (e.g., SURREAL Varol et al. (2017)). The present paper makes the following contributions.

The concept of pre-training without natural images provides a method by which to automatically generate a large-scale image dataset complete with image patterns and their labels. In order to construct such a database, through exploration research, we experimentally disclose ways to automatically generate categories using fractals. The present paper proposes two sets of randomly searched fractal databases generated in such a manner: FractalDB-1k/10k, which consists of 1,000/10,000 categories (see the supplementary material for all FractalDB-1k categories). See Figure 1(a) for Formula-driven Supervised Learning from categories of FractalDB-1k. Regarding the proposed database, the FractalDB pre-trained model outperforms some models pre-trained by human annotated datasets (see Table 6 for details). Furthermore, Figure 1(b) shows that FractalDB pre-training accelerated the convergence speed, which was much better than training from scratch and similar to ImageNet pre-training.

2 Related work

Pre-training on Large-scale Datasets. A number of large-scale datasets have been released for exploring how to extract an image representation, e.g., image classification Deng et al. (2009); Zhou et al. (2017), object detection Everingham et al. (2015); Lin et al. (2014); Krasin et al. (2017), and video classification Kay et al. (2017); Monfort et al. (2019)

. These datasets have contributed to improving the accuracy of DNNs when used as (pre-)training. Historically, in multiple aspects of evaluation, the ImageNet pre-trained model has been proved to be strong in transfer learning 

Donahue et al. (2014); Huh et al. (2016); Kornblith et al. (2019). Moreover, several larger-scale datasets have been proposed, e.g., JFT-300M Sun et al. (2017) and IG-3.5B Mahajan et al. (2018), for further improving the pre-training performance.

We are simply motivated to find a method to automatically generate a pre-training dataset without any natural images for acquiring a learning representation on image datasets. We believe that the proposed concept of pre-training without natural images will surpass the methods mentioned above in terms of fairness, privacy-violated, and ethics-related labels, in addition to the burdens of human annotation and image download.

Learning Frameworks. Supervised learning with well-studied architectures is currently the most promising framework for obtaining strong image representations Krizhevsky et al. (2012); Simonyan and Zisserman (2015); Szegedy et al. (2015); He et al. (2016); Xie et al. (2017); Howard et al. (2017); Sandler et al. (2018); Howard et al. (2019). Recently, the research community has been considering how to decrease the volume of labeled data with {un, weak, self}-supervised learning in order to avoid human labeling. In particular, self-supervised learning can be used to create a pre-trained model in a cost-efficient manner by using obvious labels. The idea is to make a simple but suitable task, called a pre-text task Doersch et al. (2015); Noroozi and Favaro (2016); Noroozi et al. (2018); Zhang et al. (2016); Noroozi et al. (2017); Gidaris et al. (2018). Though the early approaches (e.g., jigsaw puzzle Noroozi and Favaro (2016), image rotation Gidaris et al. (2018)

, and colorization 

Zhang et al. (2016)) were far from an alternative to human annotation, the more recent approaches (e.g., DeepCluster Caron et al. (2018), MoCo He et al. (2020), and SimCLR Chen et al. (2020)) are becoming closer to a human-based supervision like ImageNet.

The proposed framework is complementary to these studies because the above learning frameworks focus on how to represent a natural image based on an existing dataset. Unlike these studies, the proposed framework enables the generation of new image patterns based on a mathematical formula in addition to training labels. The SSL framework can replace the manual labeling supervised by human knowledge, however, there still exists the burdens of image downloading, privacy violations and unfair outputs.

Mathematical formula for image projection. One of the best-known formula-driven image projections is fractals. Fractal theory has been discussed in a long period (e.g., Mandelbrot (1983); Landini et al. (1995); Smith et al. (1996)). Fractal theory has been applied to rendering a graphical pattern in a simple equation Barnsley (1988); Monro and Budbridge (1995); Chen and Bi (1997) and constructing visual recognition models Pentland (1984); Varma and Garg (2007); Xu et al. (2009); Larsson et al. (2017). Although a rendered fractal pattern loses its infinite potential for representation by projection to a 2D-surface, a human can recognize the rendered fractal patterns as natural objects.

Since the success of these studies relies on the fractal geometry of naturally occurring phenomena Mandelbrot (1983); Falconer (2004), our assumption that fractals can assist learning image representations for recognizing natural scenes and objects is supported. Other methods, namely, the Bezier curve Farin (1993) and Perlin noise Perlin (2002), have also been discussed in terms of computational rendering. We also implement and compare these methods in the experimental section (see Table 9).

3 Automatically generated large-scale dataset

Figure 2: Overview of the proposed framework. Generating FractalDB: Pairs of an image and its fractal category are generated without human labeling and image downloading. Application to transfer learning: A FractalDB pre-trained convolutional network is assigned to conduct transfer learning for other datasets.

Figure 2 presents an overview of the Fractal DataBase (FractalDB), which consists of an infinite number of pairs of fractal images and their fractal categories with iterated function system (IFS) Barnsley (1988). We chose fractal geometry because the function enables to render complex patterns with a simple equation that are closely related to natural objects. All fractal categories are randomly searched (see Figure 1(a)), and the intra-category instances are expansively generated by considering category configurations such as rotation and patch. (The augmentation is shown as in Figure 2.)

In order to make a pre-trained CNN model, the FractalDB is applied to each training of the parameter optimization as follows. (i) Fractal images with paired labels are randomly sampled by a mini batch . (ii) Calculate the gradient of

to reduce the loss. (iii) Update the parameters. Note that we replace the pre-training step, such as the ImageNet pre-trained model. We also conduct the fine-tuning step as well as plain transfer learning (e.g., ImageNet pre-training and CIFAR-10 fine-tuning).

3.1 Fractal image generation

In order to construct fractals, we use IFS Barnsley (1988). In fractal analysis, an IFS is defined on a complete metric space by


where are transformation functions,

are probabilities having the sum of 1, and

is the number of transformations.

Using the IFS, a fractal is constructed by the random iteration algorithm Barnsley (1988), which repeats the following two steps for from an initial point . (i) Select a transformation from with pre-defined probabilities to determine the -th transformation. (ii) Produce a new point .

Since the focus herein is on representation learning for image recognition, we construct fractals in the 2D Euclidean space . In this case, each transformation is assumed in practice to be an affine transformation  Barnsley (1988), which has a set of six parameters for rotation and shifting:


An image representation of the fractal is obtained by drawing dots on a black background. The details of this step with its adaptable parameters is explained in Section 3.3.

3.2 Fractal categories

Undoubtedly, automatically generating categories for pre-training of image classification is a challenging task. Here, we associate the categories with fractal parameters . As shown in the experimental section, we successfully generate a number of pre-trained categories on FractalDB (see Figure 5) through formula-driven image projection by an IFS.

Since an IFS is characterized by a set of parameters and their corresponding probabilities, i.e., , we assume that a fractal category has a fixed and propose 1,000 or 10,000 randomly searched fractal categories (FractalDB-1k/10k). The reason for 1,000 categories is closely related to the experimental result for various #categories in Figure 4.

FractalDB-1k/10k consists of 1,000/10,000 different fractals (examples shown in Figure 1(a)), the parameters of which are automatically generated by repeating the following procedure. First,

is sampled from a discrete uniform distribution,

. Second, the parameter for the affine transformation is sampled from the uniform distribution on for . Third, is set to where is a rotation matrix of the affine transformation. Finally, is accepted as a new category if the filling rate of the representative image of its fractal is investigated in the experiment (see Table 5). The filling rate is calculated as the number of pixels of the fractal with respect to the total number of pixels of the image.

Figure 3: Intra-category augmentation of a leaf fractal. Here, , , , and are for rotation, and and are for shifting.

3.3 Adaptable parameters for FractalDB

As described in the experimental section, we investigated the several parameters related to fractal parameters and image rendering. The types of parameters are listed as follows.

#category and #instance. We believe that the effects of #instance on intra-category are the most effective in the pre-training task. First, we change the two parameters from 16 to 1,000 as {16, 32, 64, 128, 256, 512, 1,000}.

Patch vs. Point. We apply a 33 patch filter to generate fractal images in addition to the rendering at each 11 point. The patch filter makes variation in the pre-training phase. We repeat the following process times. We set a pixel , and then a random dot(s) with a 33 patch is inserted in the sampled area.

Filling rate . We set the filling rate from 0.05 (5%) to 0.25 (25% at 5% intervals, namely, {0.05, 0.10, 0.15, 0.20, 0.25}. Note that we could not get any randomized category at a filling rate of over 30%.

Weight of intra-category fractals (). In order to generate an intra-category image, the parameters for an image representation are varied. Intra-category images are generated by changing one of the parameters , and with weighting parameter . The basic parameter is from 0.8 to 1.2 at intervals of 0.1, i.e., {0.8, 0.9, 1.0, 1.1, 1.2}). Figure 3 shows an example of the intra-category variation in fractal images. We believe that various intra-category images help to improve the representation for image recognition.

#Dot () and image size (, ). We vary the parameters as {100K, 200K, 400K, 800K} and ( and ) as {256, 362, 512, 764, 1024}. The averaged parameter fixed as the grayscale means that the pixel value is (, , ) = (127, 127, 127) (in the case in which the pixel values are 0 to 255).

4 Experiments

In a set of experiments, we investigated the effectiveness of FractalDB and how to construct categories with the effects of configuration, as mentioned in Section 3.3. We then quantitatively evaluated and compared the proposed framework with Supervised Learning (ImageNet-1k and Places365, namely ImageNet Deng et al. (2009) and Places Zhou et al. (2017) pre-trained models) and Self-supervised Learning (Deep Cluster-10k Caron et al. (2018)) on several datasets Krizhevsky (2009); Deng et al. (2009); Zhou et al. (2017); Everingham et al. (2015); Tenenbaum (2015).

(a) CIFAR10
(b) CIFAR100
(c) ImageNet100
(d) Places30
Figure 4: Effects of #category and #instance on the CIFAR-10/100, ImageNet100 and Places30 datasets. The other parameter is fixed at 1,000, e.g. #Category is fixed at 1,000 when #Instance changed by {16, 32, 64, 128, 256, 512, 1,000}.

In order to confirm the properties of FractalDB and compare our pre-trained feature with previous studies, we used the ResNet-50. We simply replaced the pre-trained phase with our FractalDB (e.g., FractalDB-1k/10k), without changing the fine-tuning step. Moreover, in the usage of fine-tuning datasets, we conducted a standard training/validation. Through pre-training and fine-tuning, we assigned the momentum stochastic gradient descent (SGD) 

Bottou (2010)

with a value 0.9, a basic batch size of 256, and initial values of the learning rate of 0.01. The learning rate was multiplied by 0.1 when the learning epoch reached 30 and 60. Training was performed up to epoch 90. Moreover, the input image size was cropped by

[pixel] from a [pixel] input image.

4.1 Exploration study

In this subsection, we explored the configuration of formula-driven image datasets regarding Fractal generation by using CIFAR-10/100 (C10, C100), ImageNet-100 (IN100), and Places-30 datasets (P30) datasets (see the supplementary material for category lists in ImageNet-100 and Places-30). The parameters corresponding to those mentioned in Section 3.3.

#category and #instance (see Figures 4(a), 4(b), 4(c) and 4(d)) Here, the larger values tend to be better. Figure 4 indicates the effects of category and instance. We investigated the parameters with {16, 32, 64, 128, 256, 512, 1,000} on both properties. At the beginning, a larger parameter in pre-training tends to improve the accuracy in fine-tuning on all the datasets. With C10/100, we can see +7.9/+16.0 increases on the performance rate from 16 to 1,000 in #category. The improvement can be confirmed, but is relatively small for the #instance per category. The rates are +5.2/+8.9 on C10/100.

Hereafter, we assigned 1,000 [category] 1,000 [instance] as a basic dataset size and tried to train 10k categories since the #category parameter is more effective in improving the performance rates.

Patch vs. point (see Table 5) Patch with 3 3 [pixel] is better. Table 5 shows the difference between patch rendering and point rendering. We can confirm that the patch rendering is better for pre-training with 92.1 vs. 87.4 (+4.7) on C10 and 72.0 vs. 66.1 (+5.9) on C100. Moreover, when comparing random patch pattern at each patch (random) to fixed patch in image rendering (fix), performance rates increased by {+0.8, +1.6, +1.1, +1.8} on {C10, C100, IN100, P30}.

C10 C100 IN100 P30
Point 87.4 66.1 73.9 73.0
Patch (random) 92.1 72.0 78.9 73.2
Patch (fix) 92.9 73.6 80.0 75.0
Table 2: Exploration: Filling rate.
C10 C100 IN100 P30
.05 91.8 72.4 80.2 74.6
.10 92.0 72.3 80.5 75.5
.15 91.7 71.6 80.2 74.3
.20 91.3 70.8 78.8 74.7
.25 91.1 63.2 72.4 74.1
Table 3: Exploration: Weights
C10 C100 IN100 P30
.1 92.1 72.0 78.9 73.2
.2 92.4 72.7 79.2 73.9
.3 92.4 72.6 79.2 74.3
.4 92.7 73.1 79.6 74.9
.5 91.8 72.1 78.9 73.5
Table 4: Exploration: #Dot.
C10 C100 IN100 P30
100k 91.3 70.8 78.8 74.7
200k 90.9 71.0 79.2 74.8
400k 90.4 70.3 80.0 74.5
Table 5: Exploration: Image size
C10 C100 IN100 P30
256 92.9 73.6 80.0 75.0
362 92.2 73.2 80.5 75.1
512 90.9 71.0 79.2 73.0
724 90.8 71.0 79.2 73.0
1024 89.6 68.6 77.5 71.9
Table 1: Exploration: Patch vs. point.

Filling rate (see Table 5) 0.10 is better, but there is no significant change with {0.05, 0.10, 0.15}. The top scores for each dataset and the parameter are 92.0, 80.5 and 75.5 with a filling rate of 0.10 on C10, IN100 and P30, respectively. Based on these results, a filling rate of 0.10 appears to be better.

Weight of intra-category fractals (see Table 5)

Interval 0.4 is the best. A larger variance of intra-category tends to perform better in pre-training. Starting from the basic parameter at intervals of 0.1 with {0.8, 0.9, 1.0, 1.1, 1.2} (see Figure 

3), we varied the intervals as 0.1, 0.2, 0.3, 0.4, and 0.5. For the case in which the interval is 0.5, we set {0.01, 0.5, 1.0, 1.5, 2.0} in order to avoid the weighting value being set as zero. A higher variance of intra-category tends to provide higher accuracy. We confirm that the accuracies varied as {92.1, 92.4, 92.4, 92.7, 91.8} on C10, where 0.4 is the highest performance rate (92.7), but 0.5 decreases the recognition rate (91.8). We used the weight value with a 0.4 interval, i.e., {0.2, 0.6, 1.0, 1.4, 1.8}.

#Dot (see Table 5) We selected 200k by considering the accuracy and rendering time. The best parameters for each configurations are 100K on C10 (91.3), 200k on C100/P30 (71.0/74.8) and 400k on IN100 (80.0). Although a larger value is suitable on IN100, a lower value tends to be better on C10, C100, and P30. For the #dot parameter, 200k is the most balanced in terms of rendering speed and accuracy.

Image size (see Table 5) 256 256 or 362 362 is better. In terms of image size, [pixel] and [pixel] have similar performances, e.g., 73.6 (256) vs. 73.2 (362) on C100. A larger size, such as , is sparse in the image plane. Therefore, the fractal image projection produces better results in the cases of [pixel] and [pixel].

Moreover, we have additionally conducted two configurations with grayscale and color FractalDB. However, the effect of the color property appears not to be strong in the pre-training phase.

4.2 Comparison to other pre-trained datasets

We compared Scratch from random parameters, Places-30/365 Zhou et al. (2017), ImageNet-100/1k (ILSVRC’12) Deng et al. (2009), and FractalDB-1k/10k in Table 6. Since our implementation is not completely the same as a representative learning configuration, we implemented the framework fairly with the same parameters and compared the proposed method (FractalDB-1k/10k) with a baseline (Scratch, DeepCluster-10k, Places-30/365, and ImageNet-100/1k).

The proposed FractalDB pre-trained model recorded several good performance rates. We respectively describe them by comparing our Formula-driven Supervised Learning with Scratch, Self-supervised and Supervised Learning.

Comparison to training from scratch. FractalDB-1k / 10k pre-trained models recorded much higher accuracies than models trained from scratch on relatively small-scale datasets (C10/100, VOC12 and OG). In case of fine-tuning on large-scale datasets (ImageNet/Places365), the effect of pre-training was relatively small. However, in fine-tuning on Places 365, the FractalDB-10k pre-trained model helped to improve the performance rate which was also higher than ImageNet-1k pre-training (FractalDB-10k 50.8 vs. ImageNet-1k 50.3).

Method Pre-train Img Type C10 C100 IN1k P365 VOC12 OG
Scratch 87.6 62.7 76.1 49.9 58.9 1.1
DC-10k Natural Self-supervision 89.9 66.9 66.2 51.5 67.5 15.2
Places-30 Natural Supervision 90.1 67.8 69.1 69.5 6.4
Places-365 Natural Supervision 94.2 76.9 71.4 78.6 10.5
ImageNet-100 Natural Supervision 91.3 70.6 49.7 72.0 12.3
ImageNet-1k Natural Supervision 96.8 84.6 50.3 85.8 17.5
FractalDB-1k Formula Formula-supervision 93.4 75.7 70.3 49.5 58.9 20.9
FractalDB-10k Formula Formula-supervision 94.1 77.3 71.5 50.8 73.6 29.2
Table 6:

Classification accuracies of the Ours (FractalDB-1k/10k), Scratch, DeepCluster-10k (DC-10k), ImageNet-100/1k and Places-30/365 pre-trained models on representative pre-training datasets. We show the types of pre-trained image (Pre-train Img; which includes {Natural Image (Natural), Formula-driven Image (Formula)}) and Supervision types (Type; which includes {Self-supervision, Supervision, Formula-supervision}). We employed CIFAR-10 (C10), CIFAR-100 (C100), ImageNet-1k (IN1k), Places-365 (P365), classfication set of Pascal VOC 2012 (VOC12) and Omniglot (OG) datasets. The

bold and underlined values show the best scores, and bold values indicate the second best scores.

Comparison to Self-supervised Learning. We assigned the DeepCluster-10k Caron et al. (2018)

to compare the automatically generated image categories. The 10k indicates the pre-training with 10k categories. We believe that the auto-annotation with DeepCluster is the most similar method to our formula-driven image dataset. The DeepCluster-10k also assigns the same category to images that has similar image patterns based on K-means clustering. Our FractalDB-1k/10k pre-trained models outperformed the DeepCluster-10k on five different datasets, e.g., FractalDB-10k 94.1 vs. DeepCluster 89.9 (C10), 77.3 vs. DeepCluster-10k 66.9 (C100). Our method is better than the DeepCluster-10k which is a self-supervised learning method to train a feature representation in image recognition.

Comparison to Supervised Learning. We compared four types of supervised pre-training (e.g., ImageNet-1k and Places-365 datasets and their limited categories ImageNet-100 and Places-30 datasets). ImageNet-100 and Places-30 are subsets of ImageNet-1k and Places-365. The numbers correspond to the number of categories. At the beginning, our FractalDB-10k surpassed the ImageNet-100/Places-30 pre-trained models at all fine-tuning datasets. The results show that our framework is more effective than the pre-training with subsets from ImageNet-1k and Places365.

We compare the supervised pre-training methods which are the most promising pre-training approach ever. Although our FractalDB-1k/10k cannot beat them at all settings, our method partially outperformed the ImageNet-1k pre-trained model on Places-365 (FractalDB-10k 50.8 vs. ImageNet-1k 50.3) and Omniglot (FractalDB-10k 29.2 vs. ImageNet-1k 17.5) and Places-365 pre-trained model on CIFAR-100 (FractalDB-10k 77.3 vs. Places-365 76.9) and ImageNet (FractalDB-10k 71.5 vs. Places-365 71.4). The ImageNet-1k pre-trained model is much better than our proposed method on fine-tuning datasets such as C100 and VOC12 since these datasets contain similar categories such as animals and tools.

Figure 5: Noise and accuracy.
Dataset Category
Mtd PT Img C10 C100 IN1k P365 VOC12 OG
DC-10k Natural 89.9 66.9 66.2 51.2 67.5 15.2
DC-10k Formula 83.1 57.0 65.3 53.4 60.4 15.3
F1k Formula 93.4 75.7 70.3 49.5 58.9 20.9
F10k Formula 94.1 77.3 71.5 50.8 73.6 29.2
Table 7: The classification accuracies of the FractalDB-1k/10k (F1k/F10k) and DeepCluster-10k (DC-10k). Mtd/PT Img means Method and Pre-trained images.

4.3 Additional experiments

We also validated the proposed framework in terms of (i) category assignment, (ii) convergence speed, (iii) freezing parameters in fine-tuning, (iv) comparison to other formula-driven image datasets, (v) recognized category analysis and (vi) visualization of first convolutional filters and attention maps.

(i) Category assignment (see Figure 5 and Table 7). At the beginning, we validated whether the optimization can be successfully performed using the proposed FractalDB. Figure 5

show the transitioned pre-training accuracies with several rates of label noise. We randomly replaced the category labels. Here, 0% and 100% noise indicate normal training and fully randomized training, respectively. According to the results on FractalDB-1k, a CNN model can successfully classify fractal images, which are defined by iterated functions. Moreover, well-defined categories with a balanced pixel rate allow optimization on FractalDB. When fully randomized labels were assigned in FractalDB training, the architecture could not correct any images and the loss value was static (the accuracies are 0% at almost times). According to the result, we confirmed that the effect of the fractal category is reliable enough to train the image patterns.

Moreover, we used the DeepCluster-10k to automatically assign categories to the FractalDB. Table 7 indicates the comparison between category assignment with DeepCluster-10k (k-means) and FractalDB-1k/10k (IFS). We confirm that the DeepCluster-10k cannot successfully assign a category to fractal images. The gaps between IFS and k-means assignments are {11.0, 20.3, 13.2} on {C10, C100, VOC12}. This obviously indicates that our formula-driven image generation through the principle of IFS and the parameters in equation (2) works well compared to the DeepCluster-10k.

Freezing layer(s) C10 C100 IN100 P30
Fine-tuning 93.4 75.7 82.7 75.9
Conv1 92.3 72.2 77.9 74.3
Conv1–2 92.0 72.0 77.5 72.9
Conv1–3 89.3 68.0 71.0 68.5
Conv1–4 82.7 56.2 55.0 58.3
Conv1–5 49.4 24.7 21.2 31.4
Table 9: Other formula-driven image datasets with a Bezier curves and Perlin noise.
Pre-training C10 C100 IN100 P30
Scratch 87.6 60.6 75.3 70.3
Bezier-144 87.6 62.5 72.7 73.5
Bezier-1024 89.7 68.1 73.0 73.6
Perlin-100 90.9 70.2 73.0 73.3
Perlin-1296 90.4 71.1 79.7 74.2
FractalDB-1k 93.4 75.7 82.7 75.9
Table 8: Freezing parameters.
Dataset Category
C100 bee, chair, keyboard, maple tree, motor cycle,
orchid, pine tree
IN100 Kerry blue terrier, marmot, giant panda, television,
dough, valley
P30 cliff, mountain, skyscrape, tundra
Table 10: Performance rates in which FractalDB was better than the ImageNet pre-trained model on C10/C100/IN100/P30 fine-tuning.
(a) ImageNet
(b) Places365
(c) Fractal-1K
(d) Fractal-10K
(e) DC-10k
(f) Heatmaps with Grad-CAM. (Left) Input image. (Center-left) Activated heatmaps with ImageNet-1k pre-trained ResNet-50. (Center) Activated heatmaps with Places-365 pre-trained ResNet-50. (Center-Right, Right) Activated heatmaps with FractalDB-1K/10k pre-trained ResNet-50.
Figure 6: Visualization results: (a)–(e) show the activation of the 1st convolutional layer on ResNet-50, and (f) illustrates attentions with Grad-CAM Selvaraju et al. (2017).

(ii) Convergence speed (see Figure 1(b)). The transitioned pre-training accuracies values in FractalDB are similar to those of ImageNet pre-trained model and much faster than scratch from random parameters (Figure 1(b)). We validated the convergence speed in fine-tuning on C10. As the result of pre-training with FractalDB-1k, we accelerated the convergence speed in fine-tuning which is similar to the ImageNet pre-trained model.

(iii) Freezing parameters in fine-tuning (see Table 9). Although full-parameter fine-tuning is better, conv1 and 2 acquired a highly accurate image representation (Table 9). Freezing the conv1 layer provided only a -1.4 (92.0 vs. 93.4) or -2.8 (72.9 vs. 75.7) decrease from fine-tuning on C10 and C100, respectively. Comparing to other results, such as those for conv1–4/5 freezing, the bottom layer tended to train a better representation.

(iv) Comparison to other formula-driven image datasets (see Table 9). At this moment, the proposed FractalDB-1k/10k are better than other formula-driven image datasets. We assigned Perlin noise Perlin (2002) and Bezier curve Farin (1993) to generate image patterns and their categories just as FractalDB made the dataset (see the supplementary material for detailed dataset creation of the Bezier curve and Perlin noise). We confirmed that Perlin noise and the Bezier curve are also beneficial in making a pre-trained model that achieved better rates than scratch training. However, the proposed FractalDB is better than these approaches (Table 9). For a fairer comparison, we cite a similar #category in the formula-driven image datasets, namely FractalDB-1k (total #image: 1M), Bezier-1024 (1.024M) and Perlin-1296 (1.296M). The significantly improved rates are +3.0 (FractalDB-1k 93.4 vs. Perlin-1296 90.4) on C10, +4.6 (FractalDB-10k 75.7 vs. Perlin-1296 71.1) on C100, +3.0 (FractalDB-1k 82.7 vs. Perlin-1296 79.7) on IN100, and +1.7 (FractalDB-1k 75.9 vs. Perlin-1296 74.2) on P30.

(v) Recognized category analysis (see Table 10). We investigated Which categories are better recognized by the FractalDB pre-trained model compared to the ImageNet pre-trained model. Table 10 shows the category names and the classification rates. The FractalDB pre-trained model tends to be better when an image contains recursive patterns (e.g., a keyboard, maple trees).

(vi) Visualization of first convolutional filters (see Figures 6(a–e)) and attention maps (see Figure 6(f)). We visualized first convolutional filters and Grad-CAM Selvaraju et al. (2017) with pre-trained ResNet-50. As seen in ImageNet-1k/Places-365/DeepCluster-10k (Figures 6(a), 6(b) and 6(e)) and FractalDB-1k/10k pre-training (Figures 6(c) and 6(d)), our pre-trained models obviously generate different feature representations from conventional natural image datasets. Based on the experimental results, we confirmed that the proposed FractalDB successfully pre-trained a CNN model without any natural images even though the convolutional basis filters are different from the natural image pre-training with ImageNet-1k/DeepCluster-10k.

The pre-trained models with Grad-CAM can generate heatmaps fine-tuned on C10 dataset. According to the center-right and right in Figure 6(f), the FractalDB-1k/10k also look at the objects.

5 Discussion and Conclusion

We achieved the framework of pre-training without natural images through formula-driven image projection based on fractals. We successfully pre-trained models on FractalDB and fine-tuned the models on several representative datasets, including CIFAR-10/100, ImageNet, Places and Pascal VOC. The performance rates were higher than those of models trained from scratch and some supervised/self-supervised learning methods. Here, we summarize our observations through exploration as follows.

Towards a better pre-trained dataset. The proposed FractalDB pre-trained model partially outperformed ImageNet-1k/Places365 pre-trained models, e.g., FractalDB-10k 77.3 vs. Places-365 76.9 on CIFAR-100, FractalDB-10k 50.8 vs. ImageNet-1k 50.3 on Places-365. If we could improve the transfer accuracy of the pre-training without natural images, then the ImageNet dataset and the pre-trained model may be replaced so as to protect fairness, preserve privacy, and decrease annotation labor. Recently, for examples, 80M Tiny Images222 and ImageNet (human-related categories)333 have been withdrawn the publicly available images.

Are fractals a good rendering formula? We are looking for better mathematically generated image patterns and their categories. We confirmed that FractalDB is better than datasets based on the Bezier curve and Perlin Noise in the context of pre-trained model (see Table 9). Moreover, the proposed FractalDB can generate a good set of categories, e.g., the fact that the training accuracy decreased depending on the label noise (see Figures 5) and the formula-driven image generation is better than DeepCluster-10k in the most cases, as a method for category assignment (see Table 7) show how the fractal categories worked well.

A different image representation from human annotated datasets. The visual patterns pre-trained by FractalDB acquire a unique feature in a different way from ImageNet-1k (see Figure 6). In the future, steerable pre-training may be available depending on the fine-tuning task. Through our experiments, we confirm that a pre-trained dataset configuration should be adjusted. We hope that the proposed pre-training framework will suit a broader range of tasks, e.g., object detection and semantic segmentation, and will become a flexibly generated pre-training dataset.


  • This work was supported by JSPS KAKENHI Grant Number JP19H01134.

  • Computational resource of AI Bridging Cloud Infrastructure (ABCI) provided by National Institute of Advanced Industrial Science and Technology (AIST) was used.


  • M. F. Barnsley (1988) Fractals Everywhere. Academic Press. New York. Cited by: §2, §3.1, §3.1, §3.1, §3.
  • L. Bottou (2010)

    Large-Scale Machine Learning with Stochastic Gradient Descent

    In 19th International Conference on Computational Statistics (COMPSTAT), pp. 177–187. Cited by: §4.
  • J. Buolamwini and T. Gebru (2018) Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. In Conference on Fairness, Accountability and Transparency (FAT), pp. 77–91. Cited by: §1.
  • M. Caron, P. Bojanowski, A. Joulin, and M. Douze (2018)

    Deep Clustering for Unsupervised Learning of Visual Features


    European Conference on Computer Vision (ECCV)

    pp. 132–149. Cited by: §2, §4.2, §4.
  • T. Chen, S. Kornblith, M. Norouzi, and G. Hinton (2020) A Simple Framework for Contrastive Learning of Visual Representations. In International Conference on Machine Learning (ICML), Cited by: §2.
  • Y. Q. Chen and G. Bi (1997) 3-D IFS fractals as real-time graphics model. Computers & Graphics 21 (3), pp. 367–370. Cited by: §2.
  • J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei (2009) ImageNet: A Large-Scale Hierarchical Image Database. In

    The IEEE International Conference on Computer Vision and Pattern Recognition (CVPR)

    pp. 248–255. Cited by: §1, §1, §2, §4.2, §4.
  • C. Doersch, A. Gupta, and A. Efros (2015) Unsupervised Visual Representation Learning by Context Prediction. In The IEEE International Conference on Computer Vision (ICCV), pp. 1422–1430. Cited by: §2.
  • J. Donahue, Y. Jia, J. Hoffman, N. Zhang, E. Tzeng, and T. Darrell (2014) DeCAF:A deep convolutional activation feature for generic visual recognition. pp. 647–655. Cited by: §2.
  • M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman (2015) The Pascal Visual Object Classes Challenge: A Retrospective. International Journal of Computer Vision (IJCV) 111 (1), pp. 98–136. Cited by: §2, §4.
  • K. Falconer (2004) Fractal geometry: mathematical foundations and applications. In John Wiley & Sons, Cited by: §2.
  • G. Farin (1993) Curves and surfaces for computer aided geometric design: A practical guide. Academic Press. Cited by: §2, §4.3.
  • C. Fellbaum (1998) WordNet: An Electronic Lexical Database. BradfordBooks. Cited by: §1.
  • S. Gidaris, P. Singh, and N. Komodakis (2018) Unsupervised Representation Learning by Predicting Image Rotations. In International Conference on Learning Representation (ICLR), Cited by: §2.
  • K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick (2020) Momentum Contrast for Unsupervised Visual Representation Learning. In The IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §2.
  • K. He, X. Zhang, S. Ren, and J. Sun (2016) Deep Residual Learning for Image Recognition. In The IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778. Cited by: §2.
  • A. G. Howard, M. Sandler, G. Chu, L.-C. Chen, B. Chen, W. Tan, Y. Zhu, R. Pang, V. Vasudevan, Q. V. Le, and H. Adam (2019) Searching for MobileNetV3. In The IEEE International Conference on Computer Vision (ICCV), pp. 1314–1324. Cited by: §2.
  • A. G. Howard, Zhu, D. Kalenichenko, W. Wang, T. Weyand, M. An-dreetto, and H. Adam (2017) MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. In arXiv pre-print arXiv:1704.04861, Cited by: §2.
  • M. Huh, P. Agrawal, and A. A. Efros (2016) What makes ImageNet good for transfer learning?. In Advances in Neural Information Processing Systems NIPS 2016 Workshop, Cited by: §2.
  • W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev, S. M., and A. Zisserman (2017) The Kinetics Human Action Video Dataset. In arXiv pre-print arXiv:1705.06950, Cited by: §2.
  • S. Kornblith, J. Shlens, and Q. V. Le (2019) Do Better ImageNet Models Transfer Better?. In The IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2661–2671. Cited by: §2.
  • I. Krasin, T. Duerig, N. Alldrin, V. Ferrari, S. Abu-El-Haija, A. Kuznetsova, H. Rom, J. Uijlings, S. Popov, M. Malloci, J. Pont-Tuset, A. Veit, S. Belongie, V. Gomes, A. Gupta, C. Sun, G. Chechik, D. Cai, Z. Feng, D. Narayanan, and K. Murphy (2017) OpenImages: A public dataset for large-scale multi-label and multi-class image classification.. Cited by: §2.
  • A. Krizhevsky, I. Sutskever, and G. E. Hinton (2012) ImageNet Classification with Deep Convolutional Neural Networks. In Advances in Neural Information Processing Systems (NIPS) 25, F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger (Eds.), pp. 1097–1105. Cited by: §2.
  • A. Krizhevsky (2009) Learning Multiple Layers of Features from Tiny Images.. Cited by: §4.
  • G. Landini, P. I. Murry, and G. P. Misson (1995) Local connected fractal dimensions and lacunarity analyses of 60 degree fluorescein angiograms. In Investigative Ophthalmology & Visual Science, pp. 2749–2755. Cited by: §2.
  • G. Larsson, M. Maire, and G. Shakhnarovich (2017) FractalNet: Ultra-Deep Neural Networks without Residuals. In International Conference on Learning Representation (ICLR), Cited by: §2.
  • T. Lin, M. Maire, S. J. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. L. Zitnick (2014)

    Microsoft COCO: common objects in context

    In European Conference on Computer Vision (ECCV), pp. 740–755. Cited by: §2.
  • D. Mahajan, R. Girshick, V. Ramanathan, K. He, M. Paluri, Y. Li, A. Bharambe, and L. v. d. Maaten (2018) Exploring the Limits of Weakly Supervised Pretraining. In European Conference on Computer Vision (ECCV), pp. 181–196. Cited by: §2.
  • B. Mandelbrot (1983) The fractal geometry of nature. American Journal of Physics 51 (3). Cited by: §1, §2, §2.
  • M. Monfort, A. Andonian, B. Zhou, K. Ramakrishnan, S. Adel Bargal, T. Yan, L. Brown, Q. Fan, D. Gutfreund, C. Vondrick, and A. Oliva (2019) Moments in Time Dataset: one million videos for event understanding. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), Cited by: §2.
  • D. M. Monro and F. Budbridge (1995) Rendering algorithms for deteministic fractals. In IEEE Computer Graphics and Its Applications, pp. 32–41. Cited by: §2.
  • M. Noroozi and P. Favaro (2016) Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles. In European Conference on Computer Vision (ECCV), Cited by: §2.
  • M. Noroozi, H. Pirsiavash, and P. Favaro (2017) Representation Learning by Learning to Count. In The IEEE International Conference on Computer Vision (ICCV), pp. 5898–5906. Cited by: §2.
  • M. Noroozi, A. Vinjimoor, P. Favaro, and H. Pirsiavash (2018) Boosting Self-Supervised Learning via Knowledge Transfer. In The IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9359–9367. Cited by: §2.
  • A. P. Pentland (1984) Fractal-based description of natural scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 6 (6), pp. 661–674. Cited by: §2.
  • K. Perlin (2002) Improving Noise. ACM Transactions on Graphics (TOG) 21 (3), pp. 681–682. Cited by: §2, §4.3.
  • M. Sandler, A. G. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen (2018) MobileNetv2: Inverted Residuals and Linear Bottlenecks. Mobile Networks for Classification, Detection and Segmentation. In arXiv pre-print arXiv:1801.04381, Cited by: §2.
  • R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra (2017) Grad-CAM: Visual Explanations From Deep Networks via Gradient-Based Localization. In The IEEE International Conference on Computer Vision (ICCV), pp. 618–626. Cited by: Figure 6, §4.3.
  • K. Simonyan and A. Zisserman (2015) Very Deep Convolutional Networks for Large-Scale Image Recognition. In International Conference on Learning Representations (ICLR), Cited by: §2.
  • T. G. Jr. Smith, L. G. D., and W. B. Marks (1996) Fractal methods and results in cellular morphology - dimentions, lacunarity and multifractals. Journal of Neuroscience Methods 69 (2), pp. 123–136. Cited by: §2.
  • C. Sun, A. Shrivastava, S. Singh, and A. Gupta (2017)

    Revisiting Unreasonable Effectiveness of Data in Deep Learning Era

    In The IEEE International Conference on Computer Vision (ICCV), pp. 843–852. Cited by: §2.
  • C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich (2015) Going Deeper with Convolutions. In The IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1–9. Cited by: §2.
  • J. B. Tenenbaum (2015) Human-level concept learning through probabilistic program induction. Science 350 (6266), pp. 1332–1338. Cited by: §4.
  • M. Varma and R. Garg (2007) Locally invariant fractal features for statistical texture classification. In The IEEE International Conference on Computer Vision (ICCV), pp. 1–8. Cited by: §2.
  • G. Varol, J. Romero, X. Martin, N. Mahmood, M. J. Black, I. Laptev, and C. Schmid (2017) Learning from Synthetic Humans. In The IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), pp. 109–117. Cited by: §1, §1.
  • S. Xie, R. Girshick, P. Dollár, Z. Tu, and K. He (2017) Aggregated Residual Transformations for Deep Neural Networks. In The IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1492–1500. Cited by: §2.
  • Y. Xu, H. Ji, and C. Fermuller (2009) Viewpoint invariant texture description using fractal analysis. International Journal of Computer Vision (IJCV) 83 (1), pp. 85–100. Cited by: §2.
  • [48] K. Yang, K. Qinami, L. Fei-Fei, J. Deng, and O. Russakovsky Towards Fairer Datasets: Filtering and Balancing the Distribution of the People Subtree in the ImageNet Hierarchy. In Conference on Fairness, Accountability and Transparency (FAT), Cited by: §1.
  • R. Zhang, P. Isola, and A. A. Efros (2016) Colorful Image Colorization. In European Conference on Computer Vision (ECCV), Cited by: §2.
  • B. Zhou, A. Lapedriza, A. Khosla, A. Oliva, and A. Torralba (2017)

    Places: A 10 million Image Database for Scene Recognition

    IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40, pp. 1452–1464. Cited by: §1, §1, §2, §4.2, §4.