CVPR 2022 Papers

Temporal action detection (TAD) is an important yet challenging task in video understanding. It aims to simultaneously predict the semantic label and the temporal interval of every action instance in an untrimmed video. Rather than end-to-end learning, most existing methods adopt a head-only learning paradigm, where the video encoder is pre-trained for action classification, and only the detection head upon the encoder is optimized for TAD. The effect of end-to-end learning is not systematically evaluated. Besides, there lacks an in-depth study on the efficiency-accuracy trade-off in end-to-end TAD. In this paper, we present an empirical study of end-to-end temporal action detection. We validate the advantage of end-to-end learning over head-only learning and observe up to 11% performance improvement. Besides, we study the effects of multiple design choices that affect the TAD performance and speed, including detection head, video encoder, and resolution of input videos. Based on the findings, we build a mid-resolution baseline detector, which achieves the state-of-the-art performance of end-to-end methods while running more than 4x faster. We hope that this paper can serve as a guide for end-to-end learning and inspire future research in this field. Code and models are available at https://github.com/xlliu7/E2E-TAD.

Face recognition has achieved significant progress in deep learning era due to the ultra-large-scale and welllabeled datasets. However, training on the outsize datasets is time-consuming and takes up a lot of hardware resource. Therefore, designing an efficient training approach is indispensable. The heavy computational and memory costs mainly result from the million-level dimensionality of thefully connected (FC) layer. To this end, we propose a novel training approach, termed Faster Face Classification (F2C), to alleviate time and cost without sacrificing the performance. This method adopts Dynamic Class Pool (DCP) for storing and updating the identities' features dynamically, which could be regarded as a substitute for the FC layer. DCP is efficiently time-saving and cost-saving, as its smaller size with the independence from the whole face identities together. We further validate the proposed F2C method across several face benchmarks and private datasets, and display comparable results, meanwhile the speed is faster than state-of-the-art FC-based methods in terms of recognition accuracy and hardware costs. Moreover, our method is further improved by a well-designed dual data loader including indentity-based and instancebased loaders, which makes it more efficient for the updating DCP parameters.

Amodal completion is a visual task that humans perform easily but which is difficult for computer vision algorithms. The aim is to segment those object boundaries which are occluded and hence invisible. This task is particularly challenging for deep neural networks because data is difficult to obtain and annotate. Therefore, we formulate amodal segmentation as an out-of-task and out-of-distribution generalization problem. Specifically, we replace the fully connected classifier in neural networks with a Bayesian generative model of the neural network features. The model is trained from non-occluded images using bounding box annotations and class labels only, but is applied to generalize out-of-task to object segmentation and to generalize out-of-distribution to segment occluded objects. We demonstrate how such Bayesian models can naturally generalize beyond the training task labels when they learn a prior that models the object's background context and shape. Moreover, by leveraging an outlier process, Bayesian models can further generalize out-of-distribution to segment partially occluded objects and to predict their amodal object boundaries. Our algorithm outperforms alternative methods that use the same supervision by a large margin, and even outperforms methods where annotated amodal segmentations are used during training, when the amount of occlusion is large. Code is publically available at https://github.com/anonymous-submission-vision/Amodal-Bayesian.

Humans have the remarkable ability to perceive objects as a whole, even when parts of them are occluded. This ability of amodal perception forms the basis of our perceptual and cognitive understanding of our world. To enable robots to reason with this capability, we formulate and propose a novel task that we name amodal panoptic segmentation. The goal of this task is to simultaneously predict the pixel-wise semantic segmentation labels of the visible regions of stuff classes and the instance segmentation labels of both the visible and occluded regions of thing classes. To facilitate research on this new task, we extend two established benchmark datasets with pixel-level amodal panoptic segmentation labels that we make publicly available as KITTI-360-APS and BDD100K-APS. We present several strong baselines, along with the amodal panoptic quality (APQ) and amodal parsing coverage (APC) metrics to quantify the performance in an interpretable manner. Furthermore, we propose the novel amodal panoptic segmentation network (APSNet), as a first step towards addressing this task by explicitly modeling the complex relationships between the occluders and occludes. Extensive experimental evaluations demonstrate that APSNet achieves state-of-the-art performance on both benchmarks and more importantly exemplifies the utility of amodal recognition. The datasets are available at http://amodal-panoptic.cs.uni-freiburg.de

Unsupervised image-to-image (I2I) translation aims to learn a domain mapping function that can preserve the semantics of the input images without paired data. However, because the underlying semantics distributions in the source and target domains are often mismatched, current distribution matching-based methods may distort the semantics when matching distributions, resulting in the inconsistency between the input and translated images, which is known as the semantics distortion problem. In this paper, we focus on the low-level I2I translation, where the structure of images is highly related to their semantics. To alleviate semantic distortions in such translation tasks without paired supervision, we propose a novel I2I translation constraint, called Structure Consistency Constraint (SCC), to promote the consistency of image structures by reducing the randomness of color transformation in the translation process. To facilitate estimation and maximization of SCC, we propose an approximate representation of mutual information called relative Squared-loss Mutual Information (rSMI) that enjoys efficient analytic solutions. Our SCC can be easily incorporated into most existing translation models. Quantitative and qualitative comparisons on a range of low-level I2I translation tasks show that translation models with SCC outperform the original models by a significant margin with little additional computational and memory costs.

In this paper, we study a challenging problem in image restoration, namely, how to develop an all-in-one method that could recover images from a variety of unknown corruption types and levels. To this end, we propose an All-in-one Image Restoration Network (AirNet) consisting of two neural modules, named Contrastive-Based Degraded Encoder (CBDE) and Degradation-Guided Restoration Network (DGRN). The major advantages of AirNet are two-fold. First, it is an all-in-one solution which could recover various degraded images in one network. Second, AirNet is free from the prior of the corruption types and levels, which just uses the observed corrupted image to perform inference. These two advantages enable AirNet to enjoy better flexibility and higher economy in real world scenarios wherein the priors on the corruptions are hard to know and the degradation will change with space and time. Extensive experimental results show the proposed method outperforms 17 image restoration baselines on four challenging datasets. The code is available at https://github.com/XLearning-SCU/2022-CVPR-AirNet.

Time-of-flight (ToF) sensors provide an image modality fueling applications across domains, including lidar in autonomous driving, robotics, and augmented reality. Conventional ToF imaging methods estimate the depth of a scene point by sending pulses of light into a scene and measuring the time of flight of the first arriving photons that are returned from the scene, the ones directly reflected from a scene surface without any temporal delay. As such, all photons following this first response are typically considered as unwanted noise, including multi-bounce and sub-surface scattering of real-world materials. While multi-bounce scene interreflections have been extensively in recent work on non-line-of-sight imaging, we investigate temporally resolved sub-surface scattering in this work. We depart from the principle of first arrival and instead propose an all-photon ToF imaging method relying on polarization changes that analyzes both first- and late-arriving photons for shape and material scene understanding. To this end, we propose a novel capture method, reflectance model, and a reconstruction algorithm that exploits the polarization state of light changes after reflection in addition to ToF information. The proposed temporal-polarimetric imaging method allows for accurate geometric and material information of the scene by utilizing all photons captured by the system, decoded by polarization cues, outperforming all tested existing methods in simulation and experimentally.

Most methods tackle zero-shot video classification by aligning visual-semantic representations within seen classes, which limits generalization to unseen classes. To enhance model generalizability, this paper presents an end-to-end framework that preserves alignment and uniformity properties for representations on both seen and unseen classes. Specifically, we formulate a supervised contrastive loss to simultaneously align visual-semantic features (i.e., alignment) and encourage the learned features to distribute uniformly (i.e., uniformity). Unlike existing methods that only consider the alignment, we propose uniformity to preserve maximal-info of existing features, which improves the probability that unobserved features fall around observed data. Further, we synthesize features of unseen classes by proposing a class generator that interpolates and extrapolates the features of seen classes. Besides, we introduce two metrics, closeness and dispersion, to quantify the two properties and serve as new measurements of model generalizability. Experiments show that our method significantly outperforms SoTA by relative improvements of 28.1% on UCF101 and 27.0% on HMDB51. Code is available.

Quantization is an efficient network compression approach to reduce the inference time. However, existing approaches ignored the distribution difference between training and testing data, thereby inducing a large quantization error in inference. To address this issue, we propose a new quantization scheme, Alignment Quantization with ADMM-based Correlation Preservation (AlignQ), which exploits the cumulative distribution function (CDF) to align the data to be i.i.d. (independently and identically distributed) for quantization error minimization. Afterward, our theoretical analysis indicates that the significant changes in data correlations after the quantization induce a large quantization error. Accordingly, we aim to preserve the relationship of data from the original space to the aligned quantization space for retaining the prediction information. We design an optimization process by leveraging the Alternating Direction Method of Multipliers (ADMM) optimization to minimize the differences in data correlations before and after the alignment and quantization. In experiments, we visualize non-i.i.d. in training and testing data in the benchmark. We further adopt domain shift data to compare AlignQ with the state-of-the-art. Experimental results show that AlignQ achieves significant performance improvements, especially in low-bit models. Code is available at https://github.com/tinganchen/AlignQ.git.

Mixup is a powerful data augmentation method that interpolates between two or more examples in the input or feature space and between the corresponding target labels. However, how to best interpolate images is not well defined. Recent mixup methods overlay or cut-and-paste two or more objects into one image, which needs care in selecting regions. Mixup has also been connected to autoencoders, because often autoencoders generate an image that continuously deforms into another. However, such images are typically of low quality. In this work, we revisit mixup from the deformation perspective and introduce AlignMixup, where we geometrically align two images in the feature space. The correspondences allow us to interpolate between two sets of features, while keeping the locations of one set. Interestingly, this retains mostly the geometry or pose of one image and the appearance or texture of the other. We also show that an autoencoder can still improve representation learning under mixup, without the classifier ever seeing decoded images. AlignMixup outperforms state-of-the-art mixup methods on five different benchmarks.

Video-and-language pre-training has shown promising improvements on various downstream tasks. Most previous methods capture cross-modal interactions with a transformer-based multimodal encoder, not fully addressing the misalignment between unimodal video and text features. Besides, learning fine-grained visual-language alignment usually requires off-the-shelf object detectors to provide object information, which is bottlenecked by the detector's limited vocabulary and expensive computation cost. We propose Align and Prompt: an efficient and effective video-and-language pre-training framework with better cross-modal alignment. First, we introduce a video-text contrastive (VTC) loss to align unimodal video-text features at the instance level, which eases the modeling of cross-modal interactions. Then, we propose a new visually-grounded pre-training task, prompting entity modeling (PEM), which aims to learn fine-grained region-entity alignment. To achieve this, we first introduce an entity prompter module, which is trained with VTC to produce the similarity between a video crop and text prompts instantiated with entity names. The PEM task then asks the model to predict the entity pseudo-labels (i.e normalized similarity scores) for randomly-selected video crops. The resulting pre-trained model achieves state-of-the-art performance on both text-video retrieval and videoQA, outperforming prior work by a substantial margin. Our code and pre-trained models are available at https://github.com/salesforce/ALPRO.

Existing symmetric contrastive learning methods suffer from collapses (complete and dimensional) or quadratic complexity of objectives. Departure from these methods which maximize mutual information of two generated views, along either instance or feature dimension, the proposed paradigm introduces intermediate variables at the feature level, and maximizes the consistency between variables and representations of each view. Specifically, the proposed intermediate variables are the nearest group of base vectors to representations. Hence, we call the proposed method ARB (Align Representations with Base). Compared with other symmetric approaches, ARB 1) does not require negative pairs, which leads the complexity of the overall objective function is in linear order, 2) reduces feature redundancy, increasing the information density of training samples, 3) is more robust to output dimension size, which outperforms previous feature-wise arts over 28% Top-1 accuracy on ImageNet-100 under low-dimension settings.

Atlas building and image registration are important tasks for medical image analysis. Once one or multiple atlases from an image population have been constructed, commonly (1) images are warped into an atlas space to study intra-subject or inter-subject variations or (2) a possibly probabilistic atlas is warped into image space to assign anatomical labels. Atlas estimation and nonparametric transformations are computationally expensive as they usually require numerical optimization. Additionally, previous approaches for atlas building often define similarity measures between a fuzzy atlas and each individual image, which may cause alignment difficulties because a fuzzy atlas does not exhibit clear anatomical structures in contrast to the individual images. This work explores using a convolutional neural network (CNN) to jointly predict the atlas and a stationary velocity field (SVF) parameterization for diffeomorphic image registration with respect to the atlas. Our approach does not require affine pre-registrations and utilizes pairwise image alignment losses to increase registration accuracy. We evaluate our model on 3D knee magnetic resonance images (MRI) from the OAI-ZIB dataset. Our results show that the proposed framework achieves better performance than other state-of-the-art image registration algorithms, allows for end-to-end training, and for fast inference at test time.

Object encoding and identification are vital for robotic tasks such as autonomous exploration, semantic scene understanding, and re-localization. Previous approaches have attempted to either track objects or generate descriptors for object identification. However, such systems are limited to a "fixed" partial object representation from a single viewpoint. In a robot exploration setup, there is a requirement for a temporally "evolving" global object representation built as the robot observes the object from multiple viewpoints. Furthermore, given the vast distribution of unknown novel objects in the real world, the object identification process must be class-agnostic. In this context, we propose a novel temporal 3D object encoding approach, dubbed AirObject, to obtain global keypoint graph-based embeddings of objects. Specifically, the global 3D object embeddings are generated using a temporal convolutional network across structural information of multiple frames obtained from a graph attention-based encoding method. We demonstrate that AirObject achieves the state-of-the-art performance for video object identification and is robust to severe occlusion, perceptual aliasing, viewpoint shift, deformation, and scale transform, outperforming the state-of-the-art single-frame and sequential descriptors. To the best of our knowledge, AirObject is one of the first temporal object encoding methods. Source code is available at https://github.com/Nik-V9/AirObject.

Affine registration is indispensable in a comprehensive medical image registration pipeline. However, only a few studies focus on fast and robust affine registration algorithms. Most of these studies utilize convolutional neural networks (CNNs) to learn joint affine and non-parametric registration, while the standalone performance of the affine subnetwork is less explored. Moreover, existing CNN-based affine registration approaches focus either on the local misalignment or the global orientation and position of the input to predict the affine transformation matrix, which are sensitive to spatial initialization and exhibit limited generalizability apart from the training dataset. In this paper, we present a fast and robust learning-based algorithm, Coarse-to-Fine Vision Transformer (C2FViT), for 3D affine medical image registration. Our method naturally leverages the global connectivity and locality of the convolutional vision transformer and the multi-resolution strategy to learn the global affine registration. We evaluate our method on 3D brain atlas registration and template-matching normalization. Comprehensive results demonstrate that our method is superior to the existing CNNs-based affine registration methods in terms of registration accuracy, robustness and generalizability while preserving the runtime advantage of the learning-based methods. The source code is available at https://github.com/cwmok/C2FViT.

Text logo design heavily relies on the creativity and expertise of professional designers, in which arranging element layouts is one of the most important procedures. However, few attention has been paid to this task which needs to take many factors (e.g., fonts, linguistics, topics, etc.) into consideration. In this paper, we propose a content-aware layout generation network which takes glyph images and their corresponding text as input and synthesizes aesthetic layouts for them automatically. Specifically, we develop a dual-discriminator module, including a sequence discriminator and an image discriminator, to evaluate both the character placing trajectories and rendered shapes of synthesized text logos, respectively. Furthermore, we fuse the information of linguistics from texts and visual semantics from glyphs to guide layout prediction, which both play important roles in professional layout design. To train and evaluate our approach, we construct a dataset named as TextLogo3K, consisting of about 3,500 text logos and their pixel-level segmentation. Experimental studies on this dataset demonstrate the effectiveness of our approach for synthesizing visually-pleasing text logos and verify its superiority against the state of the art.

Black-box adversarial attack has aroused much research attention for its difficulty on nearly no available information of the attacked model and the additional constraint on the query budget. A common way to improve attack efficiency is to transfer the gradient information of a white-box substitute model trained on an extra dataset. In this paper, we deal with a more practical setting where a pre-trained white-box model with network parameters is provided without extra training data. To solve the model mismatch problem between the white-box and black-box models, we propose a novel algorithm EigenBA by systematically integrating gradient-based white-box method and zeroth-order optimization in black-box methods. We theoretically show the optimal directions of perturbations for each step are closely related to the right singular vectors of the Jacobian matrix of the pretrained white-box model. Extensive experiments on ImageNet, CIFAR-10 and WebVision show that EigenBA can consistently and significantly outperform state-of-the-art baselines in terms of success rate and attack efficiency.

We study joint video and language (VL) pre-training to enable cross-modality learning and benefit plentiful downstream VL tasks. Existing works either extract low-quality video features or learn limited text embedding, while neglecting that high-resolution videos and diversified semantics can significantly improve cross-modality learning. In this paper, we propose a novel High-resolution and Diversified VIdeo-LAnguage pre-training model (HD-VILA) for many visual tasks. In particular, we collect a large dataset with two distinct properties: 1) the first high-resolution dataset including 371.5k hours of 720p videos, and 2) the most diversified dataset covering 15 popular YouTube categories. To enable VL pre-training, we jointly optimize the HD-VILA model by a hybrid Transformer that learns rich spatiotemporal features, and a multimodal Transformer that enforces interactions of the learned video features with diversified texts. Our pre-training model achieves new state-of-the-art results in 10 VL understanding tasks and 2 more novel text-to-visual generation tasks. For example, we outperform SOTA models with relative increases of 40.4% R@1 in zero-shot MSR-VTT text-to-video retrieval task, and 55.4% in high-resolution dataset LSMDC. The learned VL embedding is also effective in generating visually pleasing and semantically relevant results in text-to-visual editing and super-resolution tasks.

Multi-Object Tracking (MOT) is most often approached in the tracking-by-detection paradigm, where object detections are associated through time. The association step naturally leads to discrete optimization problems. As these optimization problems are often NP-hard, they can only be solved exactly for small instances on current hardware. Adiabatic quantum computing (AQC) offers a solution for this, as it has the potential to provide a considerable speedup on a range of NP-hard optimization problems in the near future. However, current MOT formulations are unsuitable for quantum computing due to their scaling properties. In this work, we therefore propose the first MOT formulation designed to be solved with AQC. We employ an Ising model that represents the quantum mechanical system implemented on the AQC. We show that our approach is competitive compared with state-of-the-art optimization-based approaches, even when using of-the-shelf integer programming solvers. Finally, we demonstrate that our MOT problem is already solvable on the current generation of real quantum computers for small examples, and analyze the properties of the measured solutions.