CVPR 2021 Papers

Joint rolling shutter correction and deblurring (RSCD) techniques are critical for the prevalent CMOS cameras. However, current approaches are still based on conventional energy optimization and are developed for static scenes. To enable learning-based approaches to address real-world RSCD problem, we contribute the first dataset, BS-RSCD, which includes both ego-motion and object-motion in dynamic scenes. Real distorted and blurry videos with corresponding ground truth are recorded simultaneously via a beam-splitter-based acquisition system. Since direct application of existing individual rolling shutter correction (RSC) or global shutter deblurring (GSD) methods on RSCD leads to undesirable results due to inherent flaws in the network architecture, we further present the first learning-based model (JCD) for RSCD. The key idea is that we adopt bi-directional warping streams for displacement compensation, while also preserving the non-warped deblurring stream for details restoration. The experimental results demonstrate that JCD achieves state-of-the-art performance on the realistic RSCD dataset (BS-RSCD) and the synthetic RSC dataset (Fastec-RS). The dataset and code are available at https://github.com/zzh-tech/RSCD.

Despite the success of machine learning applications in science, industry, and society in general, many approaches are known to be non-robust, often relying on spurious correlations to make predictions. Spuriousness occurs when some features correlate with labels but are not causal; relying on such features prevents models from generalizing to unseen environments where such correlations break. In this work, we focus on image classification and propose two data generation processes to reduce spuriousness. Given human annotations of the subset of the features responsible (causal) for the labels (e.g. bounding boxes), we modify this causal set to generate a surrogate image that no longer has the same label (i.e. a counterfactual image). We also alter non-causal features to generate images still recognized as the original labels, which helps to learn a model invariant to these features. In several challenging datasets, our data generations outperform state-of-the-art methods in accuracy when spurious correlations break, and increase the saliency focus on causal features providing better explanations.

Blind face restoration usually relies on facial priors, such as facial geometry prior or reference prior, to restore realistic and faithful details. However, very low-quality inputs cannot offer accurate geometric prior while high-quality references are inaccessible, limiting the applicability in real-world scenarios. In this work, we propose GFP-GAN that leverages rich and diverse priors encapsulated in a pretrained face GAN for blind face restoration. This Generative Facial Prior (GFP) is incorporated into the face restoration process via spatial feature transform layers, which allow our method to achieve a good balance of realness and fidelity. Thanks to the powerful generative facial prior and delicate designs, our GFP-GAN could jointly restore facial details and enhance colors with just a single forward pass, while GAN inversion methods require image-specific optimization at inference. Extensive experiments show that our method achieves superior performance to prior art on both synthetic and real-world datasets.

Recent advances in 3D semantic scene understanding have shown impressive progress in 3D instance segmentation, enabling object-level reasoning about 3D scenes; however, a finer-grained understanding is required to enable interactions with objects and their functional understanding. Thus, we propose the task of part-based scene understanding of real-world 3D environments: from an RGB-D scan of a scene, we detect objects, and for each object predict its decomposition into geometric part masks, which composed together form the complete geometry of the observed object. We leverage an intermediary part graph representation to enable robust completion as well as building of part priors, which we use to construct the final part mask predictions. Our experiments demonstrate that guiding part understanding through part graph to part prior-based predictions significantly outperforms alternative approaches to the task of part-based instance completion.

Humans have a natural instinct to identify unknown object instances in their environments. The intrinsic curiosity about these unknown instances aids in learning about them, when the corresponding knowledge is eventually available. This motivates us to propose a novel computer vision problem called: `Open World Object Detection', where a model is tasked to: 1) identify objects that have not been introduced to it as `unknown', without explicit supervision to do so, and 2) incrementally learn these identified unknown categories without forgetting previously learned classes, when the corresponding labels are progressively received. We formulate the problem, introduce a strong evaluation protocol and provide a novel solution, which we call OREO: Open World Object Detector, based on contrastive clustering and energy based unknown identification. Our experimental evaluation and ablation studies analyse the efficacy of OREO in achieving Open World objectives. As an interesting by-product, we find that identifying and characterising unknown instances helps to reduce confusion in an incremental object detection setting, where we achieve state-of-the-art performance, with no extra methodological effort. We hope that our work will attract further research into this newly identified, yet crucial research direction.

Tracking by natural language specification is a new rising research topic that aims at locating the target object in the video sequence based on its language description. Compared with traditional bounding box (BBox) based tracking, this setting guides object tracking with high-level semantic information, addresses the ambiguity of BBox, and links local and global search organically together. Those benefits may bring more flexible, robust and accurate tracking performance in practical scenarios. However, existing natural language initialized trackers are developed and compared on benchmark datasets proposed for tracking-by-BBox, which can't reflect the true power of tracking-by-language. In this work, we propose a new benchmark specifically dedicated to the tracking-by-language, including a large scale dataset, strong and diverse baseline methods. Specifically, we collect 2k video sequences (contains a total of 1,244,340 frames, 663 words) and split 1300/700 for the train/testing respectively. We densely annotate one sentence in English and corresponding bounding boxes of the target object for each video. We also introduce two new challenges into TNL2K for the object tracking task, i.e., adversarial samples and modality switch. A strong baseline method based on an adaptive local-global-search scheme is proposed for future works to compare. We believe this benchmark will greatly boost related researches on natural language guided tracking.

Our world offers a never-ending stream of visual stimuli, yet today's vision systems only accurately recognize patterns within a few seconds. These systems understand the present, but fail to contextualize it in past or future events. In this paper, we study long-form video understanding. We introduce a framework for modeling long-form videos and develop evaluation protocols on large-scale datasets. We show that existing state-of-the-art short-term models are limited for long-form tasks. A novel object-centric transformer-based video recognition architecture performs significantly better on 7 diverse tasks. It also outperforms comparable state-of-the-art on the AVA dataset.

Most differentiable neural architecture search methods construct a super-net for search and derive a target-net as its sub-graph for evaluation. There exists a significant gap between the architectures in search and evaluation. As a result, current methods suffer from an inconsistent, inefficient, and inflexible search process. In this paper, we introduce EnTranNAS that is composed of Engine-cells and Transit-cells. The Engine-cell is differentiable for architecture search, while the Transit-cell only transits a sub-graph by architecture derivation. Consequently, the gap between the architectures in search and evaluation is significantly reduced. Our method also spares much memory and computation cost, which speeds up the search process. A feature sharing strategy is introduced for more balanced optimization and more efficient search. Furthermore, we develop an architecture derivation method to replace the traditional one that is based on a hand-crafted rule. Our method enables differentiable sparsification, and keeps the derived architecture equivalent to that of Engine-cell, which further improves the consistency between search and evaluation. More importantly, it supports the search for topology where a node can be connected to prior nodes with any number of connections, so that the searched architectures could be more flexible. Our search on CIFAR-10 has an error rate of 2.22% with only 0.07 GPU-day. We can also directly perform the search on ImageNet with topology learnable and achieve a top-1 error rate of 23.8% in 2.1 GPU-day.

Existing face relighting methods often struggle with two problems: maintaining the local facial details of the subject and accurately removing and synthesizing shadows in the relit image, especially hard shadows. We propose a novel deep face relighting method that addresses both problems. Our method learns to predict the ratio (quotient) image between a source image and the target image with the desired lighting, allowing us to relight the image while maintaining the local facial details. During training, our model also learns to accurately modify shadows by using estimated shadow masks to emphasize on the high-contrast shadow borders. Furthermore, we introduce a method to use the shadow mask to estimate the ambient light intensity in an image, and are thus able to leverage multiple datasets during training with different global lighting intensities. With quantitative and qualitative evaluations on the Multi-PIE and FFHQ datasets, we demonstrate that our proposed method faithfully maintains the local facial details of the subject and can accurately handle hard shadows while achieving state-of-the-art face relighting performance.

Data is the engine of modern computer vision, which necessitates collecting large-scale datasets. This is expensive, and guaranteeing the quality of the labels is a major challenge. In this paper, we investigate efficient annotation strategies for collecting multi-class classification labels for a large collection of images. While methods that exploit learnt models for labeling exist, a surprisingly prevalent approach is to query humans for a fixed number of labels per datum and aggregate them, which is expensive. Building on prior work on online joint probabilistic modeling of human annotations and machine-generated beliefs, we propose modifications and best practices aimed at minimizing human labeling effort. Specifically, we make use of advances in self-supervised learning, view annotation as a semi-supervised learning problem, identify and mitigate pitfalls and ablate several key design choices to propose effective guidelines for labeling. Our analysis is done in a more realistic simulation that involves querying human labelers, which uncovers issues with evaluation using existing worker simulation methods. Simulated experiments on a 125k image subset of the ImageNet100 show that it can be annotated to 80% top-1 accuracy with 0.35 annotations per image on average, a 2.7x and 6.7x improvement over prior work and manual annotation, respectively.

Depth maps obtained by commercial depth sensors are always in low-resolution, making it difficult to be used in various computer vision tasks. Thus, depth map super-resolution (SR) is a practical and valuable task, which upscales the depth map into high-resolution (HR) space. However, limited by the lack of real-world paired low-resolution (LR) and HR depth maps, most existing methods use downsampling to obtain paired training samples. To this end, we first construct a large-scale dataset named "RGB-D-D", which can greatly promote the study of depth map SR and even more depth-related real-world tasks. The "D-D" in our dataset represents the paired LR and HR depth maps captured from mobile phone and Lucid Helios respectively ranging from indoor scenes to challenging outdoor scenes. Besides, we provide a fast depth map super-resolution (FDSR) baseline, in which the high-frequency component adaptively decomposed from RGB image to guide the depth map SR. Extensive experiments on existing public datasets demonstrate the effectiveness and efficiency of our network compared with the state-of-the-art methods. Moreover, for the real-world LR depth maps, our algorithm can produce more accurate HR depth maps with clearer boundaries and to some extent correct the depth value errors.

Recurrent Neural Networks (RNNs) have been widely used in sequence analysis and modeling. However, when processing high-dimensional data, RNNs typically require very large model sizes, thereby bringing a series of deployment challenges. Although various prior works have been proposed to reduce the RNN model sizes, executing RNN models in resource-restricted environments is still a very challenging problem. In this paper, we propose to develop extremely compact RNN models with fully decomposed hierarchical Tucker (FDHT) structure. The HT decomposition does not only provide much higher storage cost reduction than the other tensor decomposition approaches but also brings better accuracy performance improvement for the compact RNN models. Meanwhile, unlike the existing tensor decomposition-based methods that can only decompose the input-to-hidden layer of RNNs, our proposed fully decomposition approach enables the comprehensive compression for the entire RNN models with maintaining very high accuracy. Our experimental results on several popular video recognition datasets show that our proposed fully decomposed hierarchical tucker-based LSTM (FDHT-LSTM) is extremely compact and highly efficient. To the best of our knowledge, FDHT-LSTM, for the first time, consistently achieves very high accuracy with only few thousand parameters (3,132 to 8,808) on different datasets. Compared with the state-of-the-art compressed RNN models, such as TT-LSTM, TR-LSTM and BT-LSTM, our FDHT-LSTM simultaneously enjoys both order-of-magnitude (3,985x to 10,711x) fewer parameters and significant accuracy improvement (0.6% to 12.7%).

Recent works have shown that interval bound propagation (IBP) can be used to train verifiably robust neural networks. Reseachers observe an intriguing phenomenon on these IBP trained networks: CROWN, a bounding method based on tight linear relaxation, often gives very loose bounds on these networks. We also observe that most neurons become dead during the IBP training process, which could hurt the representation capability of the network. In this paper, we study the relationship between IBP and CROWN, and prove that CROWN is always tighter than IBP when choosing appropriate bounding lines. We further propose a relaxed version of CROWN, linear bound propagation (LBP), that can be used to verify large networks to obtain lower verified errors than IBP. We also design a new activation function, parameterized ramp function (ParamRamp), which has more diversity of neuron status than ReLU. We conduct extensive experiments on MNIST, CIFAR-10 and Tiny-ImageNet with ParamRamp activation and achieve state-of-the-art verified robustness. Code is available at https://github.com/ZhaoyangLyu/VerifiablyRobustNN.

Advanced tensor decomposition, such as Tensor train (TT) and Tensor ring (TR), has been widely studied for deep neural network (DNN) model compression, especially for recurrent neural networks (RNNs). However, compressing convolutional neural networks (CNNs) using TT/TR always suffers significant accuracy loss. In this paper, we propose a systematic framework for tensor decomposition-based model compression using Alternating Direction Method of Multipliers (ADMM). By formulating TT decomposition-based model compression to an optimization problem with constraints on tensor ranks, we leverage ADMM technique to systemically solve this optimization problem in an iterative way. During this procedure, the entire DNN model is trained in the original structure instead of TT format, but gradually enjoys the desired low tensor rank characteristics. We then decompose this uncompressed model to TT format and fine-tune it to finally obtain a high-accuracy TT-format DNN model. Our framework is very general, and it works for both CNNs and RNNs, and can be easily modified to fit other tensor decomposition approaches. We evaluate our proposed framework on different DNN models for image classification and video recognition tasks. Experimental results show that our ADMM-based TT-format models demonstrate very high compression performance with high accuracy. Notably, on CIFAR-100, with 2.3X and 2.4X compression ratios, our models have 1.96% and 2.21% higher top-1 accuracy than the original ResNet-20 and ResNet-32, respectively. For compressing ResNet-18 on ImageNet, our model achieves 2.47X FLOPs reduction without accuracy loss.

Video paragraph captioning aims to describe multiple events in untrimmed videos with descriptive paragraphs. Existing approaches mainly solve the problem in two steps: event detection and then event captioning. Such two-step manner makes the quality of generated paragraphs highly dependent on the accuracy of event proposal detection which is already a challenging task. In this paper, we propose a paragraph captioning model which eschews the problematic event detection stage and directly generates paragraphs for untrimmed videos. To describe coherent and diverse events, we propose to enhance the conventional temporal attention with dynamic video memories, which progressively exposes new video features and suppresses over-accessed video contents to control visual focuses of the model. In addition, a diversity-driven training strategy is proposed to improve diversity of paragraph on the language perspective. Considering that untrimmed videos generally contain massive but redundant frames, we further augment the video encoder with keyframe awareness to improve efficiency. Experimental results on the ActivityNet and Charades datasets show that our proposed model significantly outperforms the state-of-the-art performance on both accuracy and diversity metrics without using any event boundary annotations. Code will be released at https://github.com/syuqings/video-paragraph.

Channel pruning and tensor decomposition have received extensive attention in convolutional neural network compression. However, these two techniques are traditionally deployed in an isolated manner, leading to significant accuracy drop when pursuing high compression rates. In this paper, we propose a Collaborative Compression (CC) scheme, which joints channel pruning and tensor decomposition to compress CNN models by simultaneously learning the model sparsity and low-rankness. Specifically, we first investigate the compression sensitivity of each layer in the network, and then propose a Global Compression Rate Optimization that transforms the decision problem of compression rate into an optimization problem. After that, we propose multi-step heuristic compression to remove redundant compression units step-by-step, which fully considers the effect of the remaining compression space (i.e., unremoved compression units). Our method demonstrates superior performance gains over previous ones on various datasets and backbone architectures. For example, we achieve 52.9% FLOPs reduction by removing 48.4% parameters on ResNet-50 with only a Top-1 accuracy drop of 0.56% on ImageNet 2012.

Dense Event Captioning (DEC) aims to jointly localize and describe multiple events of interest in untrimmed videos, which is an advancement of the conventional video captioning task (generating a single sentence description for a trimmed video). Weakly Supervised Dense Event Captioning (WS-DEC) goes one step further by not relying on human-annotated temporal event boundaries. However, there are few methods trying to tackle this task, and how to connect localization and description remains an open problem. In this paper, we demonstrate that under weak supervision, the event captioning module and localization module should be more closely bridged in order to improve description performance. Different from previous approaches, in our method, the event captioner generates a sentence from a video segment and feeds it to the sentence localizer to reconstruct the segment, and the localizer produces word importance weights as a guidance for the captioner to improve event description. To further bridge the sentence localizer and event captioner, a concept learner is adopted as the basis of the sentence localizer, which can be utilized to construct an induced set of concept features to enhance video features and improve the event captioner. Finally, our proposed method outperforms state-of-the-art WS-DEC methods on the ActivityNet Captions dataset.

Text-based image captioning (TextCap) which aims to read and reason images with texts is crucial for a machine to understand a detailed and complex scene environment, considering that texts are omnipresent in daily life. This task, however, is very challenging because an image often contains complex texts and visual information that is hard to be described comprehensively. Existing methods attempt to extend the traditional image captioning methods to solve this task, which focus on describing the overall scene of images by one global caption. This is infeasible because the complex text and visual information cannot be described well within one caption. To resolve this difficulty, we seek to generate multiple captions that accurately describe different parts of an image in detail. To achieve this purpose, there are three key challenges: 1) it is hard to decide which parts of the texts of images to copy or paraphrase; 2) it is non-trivial to capture the complex relationship between diverse texts in an image; 3) how to generate multiple captions with diverse content is still an open problem. To conquer these, we propose a novel Anchor-Captioner method. Specifically, we first find the important tokens which are supposed to be paid more attention to and consider them as anchors. Then, for each chosen anchor, we group its relevant texts to construct the corresponding anchor-centred graph (ACG). Last, based on different ACGs, we conduct the multi-view caption generation to improve the content diversity of generated captions. Experimental results show that our method not only achieves SOTA performance but also generates diverse captions to describe images.

Predicting accurate and realistic future human poses from historically observed sequences is a fundamental task in the intersection of computer vision, graphics, and artificial intelligence. Recently, continuous efforts have been devoted to addressing this issue, which has achieved remarkable progress. However, the existing work is seriously limited by complete observation, that is, once the historical motion sequence is incomplete (with missing values), it can only produce unexpected predictions or even deformities. Furthermore, due to inevitable reasons such as occlusion and the lack of equipment precision, the incompleteness of motion data occurs frequently, which hinders the practical application of current algorithms. In this work, we first notice this challenging problem, i.e., how to generate high-fidelity human motion predictions from incomplete observations. To solve it, we propose a novel multi-task graph convolutional network (MT-GCN). Specifically, the model involves two branches, in which the primary task is to focus on forecasting future 3D human actions accurately, while the auxiliary one is to repair the missing value of the incomplete observation. Both of them are integrated into a unified framework to share the spatio-temporal representation, which improves the final performance of each collaboratively. On three large-scale datasets, for various data missing scenarios in the real world, extensive experiments demonstrate that our approach is consistently superior to the state-of-the-art methods in which the missing values from incomplete observations are not explicitly analyzed.