The key challenge of sequence representation learning is to capture the long-range temporal dependencies. Typical methods for supervised sequence representation learning are built upon recurrent neural networks to capture temporal dependencies. One potential limitation of these methods is that they only model one-order information interactions explicitly between adjacent time steps in a sequence, hence the high-order interactions between nonadjacent time steps are not fully exploited. It greatly limits the capability of modeling the long-range temporal dependencies since the temporal features learned by one-order interactions cannot be maintained for a long term due to temporal information dilution and gradient vanishing. To tackle this limitation, we propose the Non-local Recurrent Neural Memory (NRNM) for supervised sequence representation learning, which performs non-local operations \MR{by means of self-attention mechanism} to learn full-order interactions within a sliding temporal memory block and models global interactions between memory blocks in a gated recurrent manner. Consequently, our model is able to capture long-range dependencies. Besides, the latent high-level features contained in high-order interactions can be distilled by our model. We validate the effectiveness and generalization of our NRNM on three types of sequence applications across different modalities, including sequence classification, step-wise sequential prediction and sequence similarity learning. Our model compares favorably against other state-of-the-art methods specifically designed for each of these sequence applications.

We present NeRF-SR, a solution for high-resolution (HR) novel view synthesis with mostly low-resolution (LR) inputs. Our method is built upon Neural Radiance Fields (NeRF) that predicts per-point density and color with a multi-layer perceptron. While producing images at arbitrary scales, NeRF struggles with resolutions that go beyond observed images. Our key insight is that NeRF has a local prior, which means predictions of a 3D point can be propagated in the nearby region and remain accurate. We first exploit it by a super-sampling strategy that shoots multiple rays at each image pixel, which enforces multi-view constraint at a sub-pixel level. Then, we show that NeRF-SR can further boost the performance of super-sampling by a refinement network that leverages the estimated depth at hand to hallucinate details from related patches on an HR reference image. Experiment results demonstrate that NeRF-SR generates high-quality results for novel view synthesis at HR on both synthetic and real-world datasets.

We propose a local adversarial disentangling network (LADN) for facial makeup and de-makeup. Central to our method are multiple and overlapping local adversarial discriminators in a content-style disentangling network for achieving local detail transfer between facial images, with the use of asymmetric loss functions for dramatic makeup styles with high-frequency details. Existing techniques do not demonstrate or fail to transfer high-frequency details in a global adversarial setting, or train a single local discriminator only to ensure image structure consistency and thus work only for relatively simple styles. Unlike others, our proposed local adversarial discriminators can distinguish whether the generated local image details are consistent with the corresponding regions in the given reference image in cross-image style transfer in an unsupervised setting. Incorporating these technical contributions, we achieve not only state-of-the-art results on conventional styles but also novel results involving complex and dramatic styles with high-frequency details covering large areas across multiple facial features. A carefully designed dataset of unpaired before and after makeup images will be released.

Current deep learning results on video generation are limited while there are only a few first results on video prediction and no relevant significant results on video completion. This is due to the severe ill-posedness inherent in these three problems. In this paper, we focus on human action videos, and propose a general, two-stage deep framework to generate human action videos with no constraints or arbitrary number of constraints, which uniformly address the three problems: video generation given no input frames, video prediction given the first few frames, and video completion given the first and last frames. To make the problem tractable, in the first stage we train a deep generative model that generates a human pose sequence from random noise. In the second stage, a skeleton-to-image network is trained, which is used to generate a human action video given the complete human pose sequence generated in the first stage. By introducing the two-stage strategy, we sidestep the original ill-posed problems while producing for the first time high-quality video generation/prediction/completion results of much longer duration. We present quantitative and qualitative evaluation to show that our two-stage approach outperforms state-of-the-art methods in video generation, prediction and video completion. Our video result demonstration can be viewed at https://iamacewhite.github.io/supp/index.html

State-of-the-art techniques in Generative Adversarial Networks (GANs) such as cycleGAN is able to learn the mapping of one image domain $X$ to another image domain $Y$ using unpaired image data. We extend the cycleGAN to ${\it Conditional}$ cycleGAN such that the mapping from $X$ to $Y$ is subjected to attribute condition $Z$. Using face image generation as an application example, where $X$ is a low resolution face image, $Y$ is a high resolution face image, and $Z$ is a set of attributes related to facial appearance (e.g. gender, hair color, smile), we present our method to incorporate $Z$ into the network, such that the hallucinated high resolution face image $Y'$ not only satisfies the low resolution constrain inherent in $X$, but also the attribute condition prescribed by $Z$. Using face feature vector extracted from face verification network as $Z$, we demonstrate the efficacy of our approach on identity-preserving face image super-resolution. Our approach is general and applicable to high-quality face image generation where specific facial attributes can be controlled easily in the automatically generated results.

Speaker identification refers to the task of localizing the face of a person who has the same identity as the ongoing voice in a video. This task not only requires collective perception over both visual and auditory signals, the robustness to handle severe quality degradations and unconstrained content variations are also indispensable. In this paper, we describe a novel multimodal Long Short-Term Memory (LSTM) architecture which seamlessly unifies both visual and auditory modalities from the beginning of each sequence input. The key idea is to extend the conventional LSTM by not only sharing weights across time steps, but also sharing weights across modalities. We show that modeling the temporal dependency across face and voice can significantly improve the robustness to content quality degradations and variations. We also found that our multimodal LSTM is robustness to distractors, namely the non-speaking identities. We applied our multimodal LSTM to The Big Bang Theory dataset and showed that our system outperforms the state-of-the-art systems in speaker identification with lower false alarm rate and higher recognition accuracy.

Robust Principal Component Analysis (RPCA) via rank minimization is a powerful tool for recovering underlying low-rank structure of clean data corrupted with sparse noise/outliers. In many low-level vision problems, not only it is known that the underlying structure of clean data is low-rank, but the exact rank of clean data is also known. Yet, when applying conventional rank minimization for those problems, the objective function is formulated in a way that does not fully utilize a priori target rank information about the problems. This observation motivates us to investigate whether there is a better alternative solution when using rank minimization. In this paper, instead of minimizing the nuclear norm, we propose to minimize the partial sum of singular values, which implicitly encourages the target rank constraint. Our experimental analyses show that, when the number of samples is deficient, our approach leads to a higher success rate than conventional rank minimization, while the solutions obtained by the two approaches are almost identical when the number of samples is more than sufficient. We apply our approach to various low-level vision problems, e.g. high dynamic range imaging, motion edge detection, photometric stereo, image alignment and recovery, and show that our results outperform those obtained by the conventional nuclear norm rank minimization method.