ICCV 2021 论文列表

Video entailment aims at determining if a hypothesis textual statement is entailed or contradicted by a premise video. The main challenge of video entailment is that it requires fine-grained reasoning to understand the complex and long story-based videos. To this end, we propose to incorporate visual grounding to the entailment by explicitly linking the entities described in the statement to the evidence in the video. If the entities are grounded in the video, we enhance the entailment judgment by focusing on the frames where the entities occur. Besides, in entailment dataset, the real/fake statements are formed in pairs with subtle discrepancy, which allows an add-on explanation module to predict which words or phrases make the statement contradictory to the video and regularize the training of the entailment judgment. Experimental results demonstrate that our approach significantly outperforms the state-of-the-art methods.

Despite the great progress of person re-identification (ReID) with the adoption of Convolutional Neural Networks, current ReID models are opaque and only outputs a scalar distance between two persons. There are few methods providing users semantically understandable explanations for why two persons are the same one or not. In this paper, we propose a post-hoc method, named Attribute-guided Metric Distillation (AMD), to explain existing ReID models. This is the first method to explore attributes to answer: 1) what and where the attributes make two persons different, and 2) how much each attribute contributes to the difference. In AMD, we design a pluggable interpreter network for target models to generate quantitative contributions of attributes and visualize accurate attention maps of the most discriminative attributes. To achieve this goal, we propose a metric distillation loss by which the interpreter learns to decompose the distance of two persons into components of attributes with knowledge distilled from the target model. Moreover, we propose an attribute prior loss to make the interpreter generate attribute-guided attention maps and to eliminate biases caused by the imbalanced distribution of attributes. This loss can guide the interpreter to focus on the exclusive and discriminative attributes rather than the large-area but common attributes of two persons. Comprehensive experiments show that the interpreter can generate effective and intuitive explanations for varied models and generalize well under cross-domain settings. As a by-product, the accuracy of target models can be further improved with our interpreter.

Have you ever looked at a painting and wondered what is the story behind it? This work presents a framework to bring art closer to people by generating comprehensive descriptions of fine-art paintings. Generating informative descriptions for artworks, however, is extremely challenging, as it requires to 1) describe multiple aspects of the image such as its style, content, or composition, and 2) provide background and contextual knowledge about the artist, their influences, or the historical period. To address these challenges, we introduce a multi-topic and knowledgeable art description framework, which modules the generated sentences according to three artistic topics and, additionally, enhances each description with external knowledge. The framework is validated through an exhaustive analysis, both quantitative and qualitative, as well as a comparative human evaluation, demonstrating outstanding results in terms of both topic diversity and information veracity.

Self-supervised methods play an increasingly important role in monocular depth estimation due to their great potential and low annotation cost. To close the gap with supervised methods, recent works take advantage of extra constraints, e.g., semantic segmentation. However, these methods will inevitably increase the burden on the model. In this paper, we show theoretical and empirical evidence that the potential capacity of self-supervised monocular depth estimation can be excavated without increasing this cost. In particular, we propose (1) a novel data augmentation approach called data grafting, which forces the model to explore more cues to infer depth besides the vertical image position, (2) an exploratory self-distillation loss, which is supervised by the self-distillation label generated by our new post-processing method - selective post-processing, and (3) the full-scale network, designed to endow the encoder with the specialization of depth estimation task and enhance the representational power of the model. Extensive experiments show that our contributions can bring significant performance improvement to the baseline with even less computational overhead, and our model, named EPCDepth, surpasses the previous state-of-the-art methods even those supervised by additional constraints.

Automation of neural architecture design has been a coveted alternative to human experts. Various search methods have been proposed aiming to find the optimal architecture in the search space. One would expect the search results to improve when the search space grows larger since it would potentially contain more performant candidates. Surprisingly, we observe that enlarging search space is unbeneficial or even detrimental to existing NAS methods such as DARTS, ProxylessNAS, and SPOS. This counterintuitive phenomenon suggests that enabling existing methods to large search space regimes is non-trivial. However, this problem is less discussed in the literature. We present a Neural Search-space Evolution (NSE) scheme, the first neural architecture search scheme designed especially for large space neural architecture search problems. The necessity of a well-designed search space with constrained size is a tacit consent in existing methods, and our NSE aims at minimizing such necessity. Specifically, the NSE starts with a search space subset, then evolves the search space by repeating two steps: 1) search an optimized space from the search space subset, 2) refill this subset from a large pool of operations that are not traversed. We further extend the flexibility of obtainable architectures by introducing a learnable multi-branch setting. With the proposed method, we achieve 77.3% top-1 retrain accuracy on ImageNet with 333M FLOPs, which yielded a state-of-the-art performance among previous auto-generated architectures that do not involve knowledge distillation or weight pruning. When the latency constraint is adopted, our result also performs better than the previous best-performing mobile models with a 77.9% Top-1 retrain accuracy. Code is available at https://github.com/orashi/NSE_NAS.

In a real-world scenario, human actions are typically out of the distribution from training data, which requires a model to both recognize the known actions and reject the unknown. Different from image data, video actions are more challenging to be recognized in an open-set setting due to the uncertain temporal dynamics and static bias of human actions. In this paper, we propose a Deep Evidential Action Recognition (DEAR) method to recognize actions in an open testing set. Specifically, we formulate the action recognition problem from the evidential deep learning (EDL) perspective and propose a novel model calibration method to regularize the EDL training. Besides, to mitigate the static bias of video representation, we propose a plug-and-play module to debias the learned representation through contrastive learning. Experimental results show that our DEAR method achieves consistent performance gain on multiple mainstream action recognition models and benchmarks. Code and pre-trained models are available at https://www.rit.edu/actionlab/dear.

3D hand pose estimation from monocular videos is a long-standing and challenging problem, which is now seeing a strong upturn. In this work, we address it for the first time using a single event camera, i.e., an asynchronous vision sensor reacting on brightness changes. Our EventHands approach has characteristics previously not demonstrated with a single RGB or depth camera such as high temporal resolution at low data throughputs and real-time performance at 1000 Hz. Due to the different data modality of event cameras compared to classical cameras, existing methods cannot be directly applied to and re-trained for event streams. We thus design a new neural approach which accepts a new event stream representation suitable for learning, which is trained on newly-generated synthetic event streams and can generalise to real data. Experiments show that EventHands outperforms recent monocular methods using a colour (or depth) camera in terms of accuracy and its ability to capture hand motions of unprecedented speed. Our method, the event stream simulator and the dataset are publicly available (see https://gvv.mpi-inf.mpg.de/projects/EventHands/).

Event camera is an emerging imaging sensor for capturing dynamics of moving objects as events, which motivates our work in estimating 3D human pose and shape from the event signals. Events, on the other hand, have their unique challenges: rather than capturing static body postures, the event signals are best at capturing local motions. This leads us to propose a two-stage deep learning approach, called EventHPE. The first-stage, FlowNet, is trained by unsupervised learning to infer optical flow from events. Both events and optical flow are closely related to human body dynamics, which are fed as input to the ShapeNet in the second stage, to estimate 3D human shapes. To mitigate the discrepancy between image-based flow (optical flow) and shape-based flow (vertices movement of human body shape), a novel flow coherence loss is introduced by exploiting the fact that both flows are originated from the identical human motion. An in-house event-based 3D human dataset is curated that comes with 3D pose and shape annotations, which is by far the largest one to our knowledge. Empirical evaluations on DHP19 dataset and our in-house dataset demonstrate the effectiveness of our approach.

Event cameras can report scene movements as an asynchronous stream of data called the events. Unlike traditional cameras, event cameras have very low latency (microseconds vs milliseconds) very high dynamic range (140dB vs 60 dB), and low power consumption, as they report changes of a scene and not a complete frame. As they re-port per pixel feature-like events and not the whole intensity frame they are immune to motion blur. However, event cameras require movement between the scene and camera to fire events ,i.e., they have no output when the scene is relatively static. Traditional cameras, however, report the whole frame of pixels at once in fixed intervals but have lower dynamic range and are prone to motion blur in case of rapid movements. We get the best from both worlds and use events and intensity images together in our complementary design and estimate dense disparity from this combination. The proposed end-to-end design combines events and images in a sequential manner and correlates them to esti-mate dense depth values. Our various experimental settings in real-world and simulated scenarios exploit the superiority of our method in predicting accurate depth values with fine details. We further extend our method to extreme cases of missing the left or right event or stereo pair and also investigate stereo depth estimation with inconsistent dynamic ranges or event thresholds on the left and right pairs

Event cameras, which output events by detecting spatio-temporal brightness changes, bring a novel paradigm to image sensors with high dynamic range and low latency. Previous works have achieved impressive performances on event-based video reconstruction by introducing convolutional neural networks (CNNs). However, intrinsic locality of convolutional operations is not capable of modeling long-range dependency, which is crucial to many vision tasks. In this paper, we present a hybrid CNN-Transformer network for event-based video reconstruction (ET-Net), which merits the fine local information from CNN and global contexts from Transformer. In addition, we further propose a Token Pyramid Aggregation strategy to implement multi-scale token integration for relating internal and intersected semantic concepts in the token-space. Experimental results demonstrate that our proposed method achieves superior performance over state-of-the-art methods on multiple real-world event datasets. The code is available at https://github.com/WarranWeng/ET-Net

Event cameras are bio-inspired sensors that respond to brightness changes asynchronously and output in the form of event streams instead of frame-based images. They own outstanding advantages compared with traditional cameras: higher temporal resolution, higher dynamic range, and lower power consumption. However, the spatial resolution of existing event cameras is insufficient and challenging to be enhanced at the hardware level while maintaining the asynchronous philosophy of circuit design. Therefore, it is imperative to explore the algorithm of event stream super-resolution, which is a non-trivial task due to the sparsity and strong spatio-temporal correlation of the events from an event camera. In this paper, we propose an end-to-end framework based on spiking neural network for event stream super-resolution, which can generate high-resolution (HR) event stream from the input low-resolution (LR) event stream. A spatiotemporal constraint learning mechanism is proposed to learn the spatial and temporal distributions of the event stream simultaneously. We validate our method on four large-scale datasets and the results show that our method achieves state-of-the-art performance. The satisfying results on two downstream applications, i.e. object classification and image reconstruction, further demonstrate the usability of our method. To prove the application potential of our method, we deploy it on a mobile platform. The high-quality HR event stream generated by our real-time system demonstrates the effectiveness and efficiency of our method.

An event camera detects the scene radiance changes and sends a sequence of asynchronous event streams with high dynamic range, high temporal resolution, and low latency. However, the spatial resolution of event cameras is limited as a trade-off for these outstanding properties. To reconstruct high-resolution intensity images from event data, we propose EvIntSR-Net that converts event data to multiple latent intensity frames to achieve super-resolution on intensity images in this paper. EvIntSR-Net bridges the domain gap between event streams and intensity frames and learns to merge a sequence of latent intensity frames in a recurrent updating manner. Experimental results show that EvIntSR-Net can reconstruct SR intensity images with higher dynamic range and fewer blurry artifacts by fusing events with intensity frames for both simulated and real-world data. Furthermore, the proposed EvIntSR-Net is able to generate high-frame-rate videos with super-resolved frames.

Surface normal estimation from a single image is an important task in 3D scene understanding. In this paper, we address two limitations shared by the existing methods: the inability to estimate the aleatoric uncertainty and lack of detail in the prediction. The proposed network estimates the per-pixel surface normal probability distribution. We introduce a new parameterization for the distribution, such that its negative log-likelihood is the angular loss with learned attenuation. The expected value of the angular error is then used as a measure of the aleatoric uncertainty. We also present a novel decoder framework where pixel-wise multi-layer perceptrons are trained on a subset of pixels sampled based on the estimated uncertainty. The proposed uncertainty-guided sampling prevents the bias in training towards large planar surfaces and improves the quality of prediction, especially near object boundaries and on small structures. Experimental results show that the proposed method outperforms the state-of-the-art in ScanNet and NYUv2, and that the estimated uncertainty correlates well with the prediction error. Code is available at https://github.com/baegwangbin/surface_normal_uncertainty.

Egocentric 3D human pose estimation using a single fisheye camera has become popular recently as it allows capturing a wide range of daily activities in unconstrained environments, which is difficult for traditional outside-in motion capture with external cameras. However, existing methods have several limitations. A prominent problem is that the estimated poses lie in the local coordinate system of the fisheye camera, rather than in the world coordinate system, which is restrictive for many applications. Furthermore, these methods suffer from limited accuracy and temporal instability due to ambiguities caused by the monocular setup and the severe occlusion in a strongly distorted egocentric perspective. To tackle these limitations, we present a new method for egocentric global 3D body pose estimation using a single head-mounted fisheye camera. To achieve accurate and temporally stable global poses, a spatio-temporal optimization is performed over a sequence of frames by minimizing heatmap reprojection errors and enforcing local and global body motion priors learned from a mocap dataset. Experimental results show that our approach outperforms state-of-the-art methods both quantitatively and qualitatively.

In various imaging problems, we only have access to compressed measurements of the underlying signals, hindering most learning-based strategies which usually require pairs of signals and associated measurements for training. Learning only from compressed measurements is impossible in general, as the compressed observations do not contain information outside the range of the forward sensing operator. We propose a new end-to-end self-supervised framework that overcomes this limitation by exploiting the equivariances present in natural signals. Our proposed learning strategy performs as well as fully supervised methods. Experiments demonstrate the potential of this framework on inverse problems including sparse-view X-ray computed tomography on real clinical data and image inpainting on natural images. Code has been made available at: https://github.com/edongdongchen/EI.

Interaction and navigation defined by natural language instructions in dynamic environments pose significant challenges for neural agents. This paper focuses on addressing two challenges: handling long sequence of subtasks, and understanding complex human instructions. We propose Episodic Transformer (E.T.), a multimodal transformer that encodes language inputs and the full episode history of visual observations and actions. To improve training, we leverage synthetic instructions as an intermediate representation that decouples understanding the visual appearance of an environment from the variations of natural language instructions. We demonstrate that encoding the history with a transformer is critical to solve compositional tasks, and that pretraining and joint training with synthetic instructions further improve the performance. Our approach sets a new state of the art on the challenging ALFRED benchmark, achieving 38.4% and 8.5% task success rates on seen and unseen test splits.

Visual understanding goes well beyond the study of images or videos on the web. To achieve complex tasks in volatile situations, the human can deeply understand the environment, quickly perceive events happening around, and continuously track objects' state changes, which are still challenging for current AI systems. To equip AI system with the ability to understand dynamic ENVironments, we build a video Question Answering dataset named Env-QA. Env-QA contains 23K egocentric videos, where each video is composed of a series of events about exploring and interacting in the environment. It also provides 85K questions to evaluate the ability of understanding the composition, layout, and state changes of the environment presented by the events in videos. Moreover, we propose a video QA model, Temporal Segmentation and Event Attention network (TSEA), which introduces event-level video representation and corresponding attention mechanisms to better extract environment information and answer questions. Comprehensive experiments demonstrate the effectiveness of our framework and show the formidable challenges of Env-QA in terms of long-term state tracking, multi-event temporal reasoning and event counting, etc.

Deep neural networks (DNNs) for the semantic segmentation of images are usually trained to operate on a predefined closed set of object classes. This is in contrast to the ""open world"" setting where DNNs are envisioned to be deployed to. From a functional safety point of view, the ability to detect so-called ""out-of-distribution"" (OoD) samples, i.e., objects outside of a DNN's semantic space, is crucial for many applications such as automated driving. A natural baseline approach to OoD detection is to threshold on the pixel-wise softmax entropy. We present a two-step procedure that significantly improves that approach. Firstly, we utilize samples from the COCO dataset as OoD proxy and introduce a second training objective to maximize the softmax entropy on these samples. Starting from pretrained semantic segmentation networks we re-train a number of DNNs on different in-distribution datasets and consistently observe improved OoD detection performance when evaluating on completely disjoint OoD datasets. Secondly, we perform a transparent post-processing step to discard false positive OoD samples by so-called ""meta classification"". To this end, we apply linear models to a set of hand-crafted metrics derived from the DNN's softmax probabilities. In our experiments we consistently observe a clear additional gain in OoD detection performance, cutting down the number of detection errors by 52% when comparing the best baseline with our results. We achieve this improvement sacrificing only marginally in original segmentation performance. Therefore, our method contributes to safer DNNs with more reliable overall system performance.

We consider the problem of Federated Learning (FL) where numerous decentralized computational nodes collaborate with each other to train a centralized machine learning model without explicitly sharing their local data samples. Such decentralized training naturally leads to issues of imbalanced or differing data distributions among the local models and challenges in fusing them into a central model. Existing FL methods deal with these issues by either sharing local parameters or fusing models via online distillation. However, such a design leads to multiple rounds of inter-node communication resulting in substantial bandwidth consumption, while also increasing the risk of data leakage and consequent privacy issues. To address these problems, we propose a new distillation-based FL framework that can preserve privacy by design, while also consuming substantially less network communication resources when compared to the current state-of-the-art. Our framework engages in inter-node communication using only publicly available and approved datasets, thereby giving explicit privacy control to the user. To distill knowledge among the various local models, our framework involves a novel ensemble distillation algorithm that uses both final prediction as well as model attention. This algorithm explicitly considers the diversity among various local nodes while also seeking consensus among them. This results in a comprehensive technique to distill knowledge from various decentralized nodes. We demonstrate the various aspects and the associated benefits of our FL framework through extensive experiments that produce state-of-the-art results on both classification and segmentation tasks on natural and medical images.