Patch-based methods and deep networks have been employed to tackle image inpainting problem, with their own strengths and weaknesses. Patch-based methods are capable of restoring a missing region with high-quality texture through searching nearest neighbor patches from the unmasked regions. However, these methods bring problematic contents when recovering large missing regions. Deep networks, on the other hand, show promising results in completing large regions. Nonetheless, the results often lack faithful and sharp details that resemble the surrounding area. By bringing together the best of both paradigms, we propose a new deep inpainting framework where texture generation is guided by a texture memory of patch samples extracted from unmasked regions. The framework has a novel design that allows texture memory retrieval to be trained end-to-end with the deep inpainting network. In addition, we introduce a patch distribution loss to encourage high-quality patch synthesis. The proposed method shows superior performance both qualitatively and quantitatively on three challenging image benchmarks, i.e., Places, CelebA-HQ, and Paris Street-View datasets.
Deep neural networks excel at finding hierarchical representations that solve complex tasks over large data sets. How can we humans understand these learned representations? In this work, we present network dissection, an analytic framework to systematically identify the semantics of individual hidden units within image classification and image generation networks. First, we analyze a convolutional neural network (CNN) trained on scene classification and discover units that match a diverse set of object concepts. We find evidence that the network has learned many object classes that play crucial roles in classifying scene classes. Second, we use a similar analytic method to analyze a generative adversarial network (GAN) model trained to generate scenes. By analyzing changes made when small sets of units are activated or deactivated, we find that objects can be added and removed from the output scenes while adapting to the context. Finally, we apply our analytic framework to understanding adversarial attacks and to semantic image editing.
Shots are key narrative elements of various videos, e.g. movies, TV series, and user-generated videos that are thriving over the Internet. The types of shots greatly influence how the underlying ideas, emotions, and messages are expressed. The technique to analyze shot types is important to the understanding of videos, which has seen increasing demand in real-world applications in this era. Classifying shot type is challenging due to the additional information required beyond the video content, such as the spatial composition of a frame and camera movement. To address these issues, we propose a learning framework Subject Guidance Network (SGNet) for shot type recognition. SGNet separates the subject and background of a shot into two streams, serving as separate guidance maps for scale and movement type classification respectively. To facilitate shot type analysis and model evaluations, we build a large-scale dataset MovieShots, which contains 46K shots from 7K movie trailers with annotations of their scale and movement types. Experiments show that our framework is able to recognize these two attributes of shot accurately, outperforming all the previous methods.
Generative Adversarial Networks (GANs) have recently advanced image synthesis by learning the underlying distribution of observed data in an unsupervised manner. However, how the features trained from solving the task of image synthesis are applicable to visual tasks remains seldom explored. In this work, we show that learning to synthesize images is able to bring remarkable hierarchical visual features that are generalizable across a wide range of visual tasks. Specifically, we consider the pre-trained StyleGAN generator as a learned loss function and utilize its layer-wise disentangled representation to train a novel hierarchical encoder. As a result, the visual feature produced by our encoder, termed as Generative Hierarchical Feature (GH-Feat), has compelling discriminative and disentangled properties, facilitating a range of both discriminative and generative tasks. Extensive experiments on face verification, landmark detection, layout prediction, transfer learning, style mixing, and image editing show the appealing performance of the GH-Feat learned from synthesizing images, outperforming existing unsupervised feature learning methods.
A rich set of semantic attributes has been shown to emerge in the latent space of the Generative Adversarial Networks (GANs) trained for synthesizing images. In order to identify such latent semantics for image manipulation, previous methods annotate a collection of synthesized samples and then train supervised classifiers in the latent space. However, they require a clear definition of the target attribute as well as the corresponding manual annotations, severely limiting their applications in practice. In this work, we examine the internal representation learned by GANs to reveal the underlying variation factors in an unsupervised manner. By studying the essential role of the fully-connected layer that takes the latent code into the generator of GANs, we propose a general closed-form factorization method for latent semantic discovery. The properties of the identified semantics are further analyzed both theoretically and empirically. With its fast and efficient implementation, our approach is capable of not only finding latent semantics as accurately as the state-of-the-art supervised methods, but also resulting in far more versatile semantic classes across multiple GAN models trained on a wide range of datasets.
Recent attempts for unsupervised landmark learning leverage synthesized image pairs that are similar in appearance but different in poses. These methods learn landmarks by encouraging the consistency between the original images and the images reconstructed from swapped appearances and poses. While synthesized image pairs are created by applying pre-defined transformations, they can not fully reflect the real variances in both appearances and poses. In this paper, we aim to open the possibility of learning landmarks on unpaired data (i.e. unaligned image pairs) sampled from a natural image collection, so that they can be different in both appearances and poses. To this end, we propose a cross-image cycle consistency framework ($C^3$) which applies the swapping-reconstruction strategy twice to obtain the final supervision. Moreover, a cross-image flow module is further introduced to impose the equivariance between estimated landmarks across images. Through comprehensive experiments, our proposed framework is shown to outperform strong baselines by a large margin. Besides quantitative results, we also provide visualization and interpretation on our learned models, which not only verifies the effectiveness of the learned landmarks, but also leads to important insights that are beneficial for future research.
Visual tempo, which describes how fast an action goes, has shown its potential in supervised action recognition. In this work, we demonstrate that visual tempo can also serve as a self-supervision signal for video representation learning. We propose to maximize the mutual information between representations of slow and fast videos via hierarchical contrastive learning (VTHCL). Specifically, by sampling the same instance at slow and fast frame rates respectively, we can obtain slow and fast video frames which share the same semantics but contain different visual tempos. Video representations learned from VTHCL achieve the competitive performances under the self-supervision evaluation protocol for action recognition on UCF-101 (82.1\%) and HMDB-51 (49.2\%). Moreover, we show that the learned representations are also generalized well to other downstream tasks including action detection on AVA and action anticipation on Epic-Kitchen. Finally, our empirical analysis suggests that a more thorough evaluation protocol is needed to verify the effectiveness of the self-supervised video representations across network structures and downstream tasks.
Conventional Reinforcement Learning (RL) algorithms usually have one single agent learning to solve the task independently. As a result, the agent can only explore a limited part of the state-action space while the learned behavior is highly correlated to the agent's previous experience, making the training prone to a local minimum. In this work, we empower RL with the capability of teamwork and propose a novel non-local policy optimization framework called Diversity-regularized Collaborative Exploration (DiCE). DiCE utilizes a group of heterogeneous agents to explore the environment simultaneously and share the collected experiences. A regularization mechanism is further designed to maintain the diversity of the team and modulate the exploration. We implement the framework in both on-policy and off-policy settings and the experimental results show that DiCE can achieve substantial improvement over the baselines in the MuJoCo locomotion tasks.
Despite the remarkable progress made by the policy gradient algorithms in reinforcement learning (RL), sub-optimal policies usually result from the local exploration property of the policy gradient update. In this work, we propose a method referred to as Zeroth-Order Supervised Policy Improvement (ZOSPI) that exploits the estimated value function Q globally while preserves the local exploitation of the policy gradient methods. We prove that with a good function structure, the zeroth-order optimization strategy combining both local and global samplings can find the global minima within a polynomial number of samples. To improve the exploration efficiency in unknown environments, ZOSPI is further combined with bootstrapped Q networks. Different from the standard policy gradient methods, the policy learning of ZOSPI is conducted in a self-supervision manner so that the policy can be implemented with gradient-free non-parametric models besides the neural network approximator. Experiments show that ZOSPI achieves competitive results on MuJoCo locomotion tasks with a remarkable sample efficiency.
In this work, we address the problem of learning to seek novel policies in reinforcement learning tasks. Instead of following the multi-objective framework used in previous methods, we propose to rethink the problem under a novel perspective of constrained optimization. We first introduce a new metric to evaluate the difference between policies, and then design two practical novel policy seeking methods following the new perspective, namely the Constrained Task Novel Bisector (CTNB), and the Interior Policy Differentiation (IPD), corresponding to the feasible direction method and the interior point method commonly known in constrained optimization problems. Experimental comparisons on the MuJuCo control suite show our methods achieve substantial improvements over previous novelty-seeking methods in terms of both novelty and primal task performance.