We propose a Paired Few-shot GAN (PFS-GAN) model for learning generators with sufficient source data and a few target data. While generative model learning typically needs large-scale training data, our PFS-GAN not only uses the concept of few-shot learning but also domain shift to transfer the knowledge across domains, which alleviates the issue of obtaining low-quality generator when only trained with target domain data. The cross-domain datasets are assumed to have two properties: (1) each target-domain sample has its source-domain correspondence and (2) two domains share similar content information but different appearance. Our PFS-GAN aims to learn the disentangled representation from images, which composed of domain-invariant content features and domain-specific appearance features. Furthermore, a relation loss is introduced on the content features while shifting the appearance features to increase the structural diversity. Extensive experiments show that our method has better quantitative and qualitative results on the generated target-domain data with higher diversity in comparison to several baselines.
In this paper, we address the problem of distillation-based class-incremental learning with a single head. A central theme of this task is to learn new classes that arrive in sequential phases over time while keeping the model's capability of recognizing seen classes with only limited memory for preserving seen data samples. Many regularization strategies have been proposed to mitigate the phenomenon of catastrophic forgetting. To understand better the essence of these regularizations, we introduce a feature-graph preservation perspective. Insights into their merits and faults motivate our weighted-Euclidean regularization for old knowledge preservation. We further propose rectified cosine normalization and show how it can work with binary cross-entropy to increase class separation for effective learning of new classes. Experimental results on both CIFAR-100 and ImageNet datasets demonstrate that our method outperforms the state-of-the-art approaches in reducing classification error, easing catastrophic forgetting, and encouraging evenly balanced accuracy over different classes. Our project page is at : https://github.com/yhchen12101/FGP-ICL.
The current neural networks for semantic segmentation usually predict the pixel-wise semantics on the down-sampled grid of images to alleviate the computational cost for dense maps. However, the accuracy of resultant segmentation maps may also be down graded particularly in the regions near object boundaries. In this paper, we advance to have a deeper investigation on the sampling efficiency of the down-sampled grid. By applying the spectral analysis that analyze on the network back propagation process in frequency domain, we discover that cross-entropy is mainly contributed by the low-frequency components of segmentation maps, as well as that of the feature in CNNs. The network performance maintains as long as the resolution of the down sampled grid meets the cut-off frequency. Such finding leads us to propose a simple yet effective feature truncation method that limits the feature size in CNNs and removes the associated high-frequency components. This method can not only reduce the computational cost but also maintain the performance of semantic segmentation networks. Moreover, one can seamlessly integrate this method with the typical network pruning approaches for further model reduction. On the other hand, we propose to employee a block-wise weak annotation for semantic segmentation that captures the low-frequency information of the segmentation map and is easy to collect. Using the proposed analysis scheme, one can easily estimate the efficacy of the block-wise annotation and the feature truncation method.
There exist many forms of deep latent variable models, such as the variational autoencoder and adversarial autoencoder. Regardless of the specific class of model, there exists an implicit consensus that the latent distribution should be regularized towards the prior, even in the case where the prior distribution is learned. Upon investigating the effect of latent regularization on image generation our results indicate that in the case where a sufficiently expressive prior is learned, latent regularization is not necessary and may in fact be harmful insofar as image quality is concerned. We additionally investigate the benefit of learned priors on two common problems in computer vision: latent variable disentanglement, and diversity in image-to-image translation.
Recently proposed normalizing flow models such as Glow have been shown to be able to generate high quality, high dimensional images with relatively fast sampling speed. Due to their inherently restrictive architecture, however, it is necessary that they are excessively deep in order to train effectively. In this paper we propose to combine Glow with an underlying variational autoencoder in order to counteract this issue. We demonstrate that our proposed model is competitive with Glow in terms of image quality and test likelihood while requiring far less time for training.
Inferring the information of 3D layout from a single equirectangular panorama is crucial for numerous applications of virtual reality or robotics (e.g., scene understanding and navigation). To achieve this, several datasets are collected for the task of 360 layout estimation. To facilitate the learning algorithms for autonomous systems in indoor scenarios, we consider the Matterport3D dataset with their originally provided depth map ground truths and further release our annotations for layout ground truths from a subset of Matterport3D. As Matterport3D contains accurate depth ground truths from time-of-flight (ToF) sensors, our dataset provides both the layout and depth information, which enables the opportunity to explore the environment by integrating both cues.
Recently, end-to-end trainable deep neural networks have significantly improved stereo depth estimation for perspective images. However, 360{\deg} images captured under equirectangular projection cannot benefit from directly adopting existing methods due to distortion introduced (i.e., lines in 3D are not projected onto lines in 2D). To tackle this issue, we present a novel architecture specifically designed for spherical disparity using the setting of top-bottom 360{\deg} camera pairs. Moreover, we propose to mitigate the distortion issue by (1) an additional input branch capturing the position and relation of each pixel in the spherical coordinate, and (2) a cost volume built upon a learnable shifting filter. Due to the lack of 360{\deg} stereo data, we collect two 360{\deg} stereo datasets from Matterport3D and Stanford3D for training and evaluation. Extensive experiments and ablation study are provided to validate our method against existing algorithms. Finally, we show promising results on real-world environments capturing images with two consumer-level cameras.
Stereo matching and flow estimation are two essential tasks for scene understanding, spatially in 3D and temporally in motion. Existing approaches have been focused on the unsupervised setting due to the limited resource to obtain the large-scale ground truth data. To construct a self-learnable objective, co-related tasks are often linked together to form a joint framework. However, the prior work usually utilizes independent networks for each task, thus not allowing to learn shared feature representations across models. In this paper, we propose a single and principled network to jointly learn spatiotemporal correspondence for stereo matching and flow estimation, with a newly designed geometric connection as the unsupervised signal for temporally adjacent stereo pairs. We show that our method performs favorably against several state-of-the-art baselines for both unsupervised depth and flow estimation on the KITTI benchmark dataset.