CVPR 2022 Papers

With the ubiquity of rolling shutter (RS) cameras, it is becoming increasingly attractive to recover the latent global shutter (GS) video from two consecutive RS frames, which also places a higher demand on realism. Existing solutions, using deep neural networks or optimization, achieve promising performance. However, these methods generate intermediate GS frames through image warping based on the RS model, which inevitably result in black holes and noticeable motion artifacts. In this paper, we alleviate these issues by proposing a context-aware GS video reconstruction architecture. It facilitates the advantages such as occlusion reasoning, motion compensation, and temporal abstraction. Specifically, we first estimate the bilateral motion field so that the pixels of the two RS frames are warped to a common GS frame accordingly. Then, a refinement scheme is proposed to guide the GS frame synthesis along with bilateral occlusion masks to produce high-fidelity GS video frames at arbitrary times. Furthermore, we derive an approximated bilateral motion field model, which can serve as an alternative to provide a simple but effective GS frame initialization for related tasks. Experiments on synthetic and real data show that our approach achieves superior performance over state-of-the-art methods in terms of objective metrics and subjective visual quality.

Temporal alignment of fine-grained human actions in videos is important for numerous applications in computer vision, robotics, and mixed reality. State-of-the-art methods directly learn image-based embedding space by leveraging powerful deep convolutional neural networks. While being straightforward, their results are far from satisfactory, the aligned videos exhibit severe temporal discontinuity without additional post-processing steps. The recent advancements in human body and hand pose estimation in the wild promise new ways of addressing the task of human action alignment in videos. In this work, based on off-the-shelf human pose estimators, we propose a novel context-aware self-supervised learning architecture to align sequences of actions. We name it CASA. Specifically, CASA employs self-attention and cross-attention mechanisms to incorporate the spatial and temporal context of human actions, which can solve the temporal discontinuity problem. Moreover, we introduce a self-supervised learning scheme that is empowered by novel 4D augmentation techniques for 3D skeleton representations. We systematically evaluate the key components of our method. Our experiments on three public datasets demonstrate CASA significantly improves phase progress and Kendall's Tau scores over the previous state-of-the-art methods.

High annotation costs are a substantial bottleneck in applying modern deep learning architectures to clinically relevant medical use cases, substantiating the need for novel algorithms to learn from unlabeled data. In this work, we propose ContIG, a self-supervised method that can learn from large datasets of unlabeled medical images and genetic data. Our approach aligns images and several genetic modalities in the feature space using a contrastive loss. We design our method to integrate multiple modalities of each individual person in the same model end-to-end, even when the available modalities vary across individuals. Our procedure outperforms state-of-the-art self-supervised methods on all evaluated downstream benchmark tasks. We also adapt gradient-based explainability algorithms to better understand the learned cross-modal associations between the images and genetic modalities. Finally, we perform genome-wide association studies on the features learned by our models, uncovering interesting relationships between images and genetic data.

Continually learning new classes from fresh data without forgetting previous knowledge of old classes is a very challenging research problem. Moreover, it is imperative that such learning must respect certain memory and computational constraints such as (i) training samples are limited to only a few per class, (ii) the computational cost of learning a novel class remains constant, and (iii) the memory footprint of the model grows at most linearly with the number of classes observed. To meet the above constraints, we propose C-FSCIL, which is architecturally composed of a frozen meta-learned feature extractor, a trainable fixed-size fully connected layer, and a rewritable dynamically growing memory that stores as many vectors as the number of encountered classes. C-FSCIL provides three update modes that offer a trade-off between accuracy and compute-memory cost of learning novel classes. C-FSCIL exploits hyperdimensional embedding that allows to continually express many more classes than the fixed dimensions in the vector space, with minimal interference. The quality of class vector representations is further improved by aligning them quasi-orthogonally to each other by means of novel loss functions. Experiments on the CIFAR100, miniImageNet, and Omniglot datasets show that C-FSCIL outperforms the baselines with remarkable accuracy and compression. It also scales up to the largest problem size ever tried in this few-shot setting by learning 423 novel classes on top of 1200 base classes with less than 1.6% accuracy drop. Our code is available at https://github.com/IBM/constrained-FSCIL

Post-hoc explanation methods, e.g., Grad-CAM, enable humans to inspect the spatial regions responsible for a particular network decision. However, it is shown that such explanations are not always consistent with human priors, such as consistency across image transformations. Given an interpretation algorithm, e.g., Grad-CAM, we introduce a novel training method to train the model to produce more consistent explanations. Since obtaining the ground truth for a desired model interpretation is not a well-defined task, we adopt ideas from contrastive self-supervised learning, and apply them to the interpretations of the model rather than its embeddings. We show that our method, Contrastive Grad-CAM Consistency (CGC), results in Grad-CAM interpretation heatmaps that are more consistent with human annotations while still achieving comparable classification accuracy. Moreover, our method acts as a regularizer and improves the accuracy on limited-data, fine-grained classification settings. In addition, because our method does not rely on annotations, it allows for the incorporation of unlabeled data into training, which enables better generalization of the model. The code is available here: https://github.com/UCDvision/CGC

Human-Object Interaction detection is a holistic visual recognition task that entails object detection as well as interaction classification. Previous works of HOI detection has been addressed by the various compositions of subset predictions, e.g., Image -> HO -> I, Image -> HI -> O. Recently, transformer based architecture for HOI has emerged, which directly predicts the HOI triplets in an end-to-end fashion (Image -> HOI). Motivated by various inference paths for HOI detection, we propose cross-path consistency learning (CPC), which is a novel end-to-end learning strategy to improve HOI detection for transformers by leveraging augmented decoding paths. CPC learning enforces all the possible predictions from permuted inference sequences to be consistent. This simple scheme makes the model learn consistent representations, thereby improving generalization without increasing model capacity. Our experiments demonstrate the effectiveness of our method, and we achieved significant improvement on V-COCO and HICO-DET compared to the baseline models. Our code is available at https://github.com/mlvlab/CPChoi.

Most hard attention models initially observe a complete scene to locate and sense informative glimpses, and predict class-label of a scene based on glimpses. However, in many applications (e.g., aerial imaging), observing an entire scene is not always feasible due to the limited time and resources available for acquisition. In this paper, we develop a Sequential Transformers Attention Model (STAM) that only partially observes a complete image and predicts informative glimpse locations solely based on past glimpses. We design our agent using DeiT-distilled and train it with a one-step actor-critic algorithm. Furthermore, to improve classification performance, we introduce a novel training objective, which enforces consistency between the class distribution predicted by a teacher model from a complete image and the class distribution predicted by our agent using glimpses. When the agent senses only 4% of the total image area, the inclusion of the proposed consistency loss in our training objective yields 3% and 8% higher accuracy on ImageNet and fMoW datasets, respectively. Moreover, our agent outperforms previous state-of-the-art by observing nearly 27% and 42% fewer pixels in glimpses on ImageNet and fMoW.

We attempt to connect the data from complementary views, i.e., top view from drone-mounted cameras in the air, and side view from wearable cameras on the ground. Collaborative analysis of such complementary-view data can facilitate to build the air-ground cooperative visual system for various kinds of applications. This is a very challenging problem due to the large view difference between top and side views. In this paper, we develop a new approach that can simultaneously handle three tasks: i) localizing the side-view camera in the top view; ii) estimating the view direction of the side-view camera; iii) detecting and associating the same subjects on the ground across the complementary views. Our main idea is to explore the spatial position layout of the subjects in two views. In particular, we propose a spatial-aware position representation method to embed the spatial-position distribution of the subjects in different views. We further design a cross-view video collaboration framework composed of a camera identification module and a subject association module to simultaneously perform the above three tasks. We collect a new synthetic dataset consisting of top-view and side-view video sequence pairs for performance evaluation and the experimental results show the effectiveness of the proposed method.

It's been a long history that most object detection methods obtain objects by using the non-maximum suppression (NMS) and its improved versions like Soft-NMS to remove redundant bounding boxes. We challenge those NMS-based methods from three aspects: 1) The bounding box with highest confidence value may not be the true positive having the biggest overlap with the ground-truth box. 2) Not only suppression is required for redundant boxes, but also confidence enhancement is needed for those true positives. 3) Sorting candidate boxes by confidence values is not necessary so that full parallelism is achievable. In this paper, inspired by belief propagation (BP), we propose the Confidence Propagation Cluster (CP-Cluster) to replace NMS-based methods, which is fully parallelizable as well as better in accuracy. In CP-Cluster, we borrow the message passing mechanism from BP to penalize redundant boxes and enhance true positives simultaneously in an iterative way until convergence. We verified the effectiveness of CP-Cluster by applying it to various mainstream detectors such as FasterRCNN, SSD, FCOS, YOLOv3, YOLOv5, Centernet etc. Experiments on MS COCO show that our plug and play method, without retraining detectors, is able to steadily improve average mAP of all those state-of-theart models with a clear margin from 0.3 to 1.9 respectively when compared with NMS-based methods.

With the rise of powerful pre-trained vision-language models like CLIP, it becomes essential to investigate ways to adapt these models to downstream datasets. A recently proposed method named Context Optimization (CoOp) introduces the concept of prompt learning---a recent trend in NLP---to the vision domain for adapting pre-trained vision-language models. Specifically, CoOp turns context words in a prompt into a set of learnable vectors and, with only a few labeled images for learning, can achieve huge improvements over intensively-tuned manual prompts. In our study we identify a critical problem of CoOp: the learned context is not generalizable to wider unseen classes within the same dataset, suggesting that CoOp overfits base classes observed during training. To address the problem, we propose Conditional Context Optimization (CoCoOp), which extends CoOp by further learning a lightweight neural network to generate for each image an input-conditional token (vector). Compared to CoOp's static prompts, our dynamic prompts adapt to each instance and are thus less sensitive to class shift. Extensive experiments show that CoCoOp generalizes much better than CoOp to unseen classes, even showing promising transferability beyond a single dataset; and yields stronger domain generalization performance as well. Code is available at https://github.com/KaiyangZhou/CoOp.

Convolutional neural networks (CNNs) rely on the depth of the architecture to obtain complex features. It results in computationally expensive models for low-resource IoT devices. Convolutional operators are local and restricted in the receptive field, which increases with depth. We explore partial differential equations (PDEs) that offer a global receptive field without the added overhead of maintaining large kernel convolutional filters. We propose a new feature layer, called the Global layer, that enforces PDE constraints on the feature maps, resulting in rich features. These constraints are solved by embedding iterative schemes in the network. The proposed layer can be embedded in any deep CNN to transform it into a shallower network. Thus, resulting in compact and computationally efficient architectures achieving similar performance as the original network. Our experimental evaluation demonstrates that architectures with global layers require 2-5xless computational and storage budget without any significant loss in performance.

Progress in 3D object understanding has relied on manually "canonicalized" shape datasets that contain instances with consistent position and orientation (3D pose). This has made it hard to generalize these methods to in-the-wild shapes, e.g., from internet model collections or depth sensors. ConDor is a self-supervised method that learns to Canonicalize the 3D orientation and position for full and partial 3D point clouds. We build on top of Tensor Field Networks (TFNs), a class of permutation- and rotation-equivariant, and translation-invariant 3D networks. During inference, our method takes an unseen full or partial 3D point cloud at an arbitrary pose and outputs an equivariant canonical pose. During training, this network uses self-supervision losses to learn the canonical pose from an un-canonicalized collection of full and partial 3D point clouds. ConDor can also learn to consistently co-segment object parts without any supervision. Extensive quantitative results on four new metrics show that our approach outperforms existing methods while enabling new applications such as operation on depth images and annotation transfer.

When the images are formulated as discrete measures, computing Wasserstein-p distance between them is challenging due to the complexity of solving the corresponding Kantorovich's problem. In this paper, we propose a novel algorithm to compute the Wasserstein-p distance between discrete measures by restricting the optimal transport (OT) problem on a subset. First, we define the restricted OT problem and prove the solution of the restricted problem converges to antorovich's OT solution. Second, we propose the SparseSinkhorn algorithm for the restricted problem and provide a multi-scale algorithm to estimate the subset. Finally, we implement the proposed algorithm on CUDA and illustrate the linear computational cost in terms of time and memory requirements. We compute Wasserstein-p distance, estimate the transport mapping, and transfer color between color images with size ranges from 64x64 to 1920x1200. (Our code is available at https://github.com/ucascnic/CudaOT)

Single-photon avalanche diodes (SPADs) are an emerging pixel technology for time-of-flight (ToF) 3D cameras that can capture the time-of-arrival of individual photons at picosecond resolution. To estimate depths, current SPAD-based 3D cameras measure the round-trip time of a laser pulse by building a per-pixel histogram of photon timestamps. As the spatial and timestamp resolution of SPAD-based cameras increase, their output data rates far exceed the capacity of existing data transfer technologies. One major reason for SPAD's bandwidth-intensive operation is the tight coupling that exists between depth resolution and histogram resolution. To weaken this coupling, we propose compressive single-photon histograms (CSPH). CSPHs are a per-pixel compressive representation of the high-resolution histogram, that is built on-the-fly, as each photon is detected. They are based on a family of linear coding schemes that can be expressed as a simple matrix operation. We design different CSPH coding schemes for 3D imaging and evaluate them under different signal and background levels, laser waveforms, and illumination setups. Our results show that a well-designed CSPH can consistently reduce data rates by 1-2 orders of magnitude without compromising depth precision.

Few-sample compression aims to compress a big redundant model into a small compact one with only few samples. If we fine-tune models with these limited few samples directly, models will be vulnerable to overfit and learn almost nothing. Hence, previous methods optimize the compressed model layer-by-layer and try to make every layer have the same outputs as the corresponding layer in the teacher model, which is cumbersome. In this paper, we propose a new framework named Mimicking then Replacing (MiR) for few-sample compression, which firstly urges the pruned model to output the same features as the teacher's in the penultimate layer, and then replaces teacher's layers before penultimate with a well-tuned compact one. Unlike previous layer-wise reconstruction methods, our MiR optimizes the entire network holistically, which is not only simple and effective, but also unsupervised and general. MiR outperforms previous methods with large margins. Codes is available at https://github.com/cjnjuwhy/MiR.

Comprehending the rich semantics in an image and ordering them in linguistic order are essential to compose a visually-grounded and linguistically coherent description for image captioning. Modern techniques commonly capitalize on a pre-trained object detector/classifier to mine the semantics in an image, while leaving the inherent linguistic ordering of semantics under-exploited. In this paper, we propose a new recipe of Transformer-style structure, namely Comprehending and Ordering Semantics Networks (COS-Net), that novelly unifies an enriched semantic comprehending and a learnable semantic ordering processes into a single architecture. Technically, we initially utilize a cross-modal retrieval model to search the relevant sentences of each image, and all words in the searched sentences are taken as primary semantic cues. Next, a novel semantic comprehender is devised to filter out the irrelevant semantic words in primary semantic cues, and meanwhile infer the missing relevant semantic words visually grounded in the image. After that, we feed all the screened and enriched semantic words into a semantic ranker, which learns to allocate all semantic words in linguistic order as humans. Such sequence of ordered semantic words are further integrated with visual tokens of images to trigger sentence generation. Empirical evidences show that COS-Net clearly surpasses the state-of-the-art approaches on COCO and achieves to-date the best CIDEr score of 141.1% on Karpathy test split. Source code is available at https://github.com/YehLi/xmodaler/tree/master/configs/image_caption/cosnet.

Domain generalization (DG) aims to improve the generalization performance for an unseen target domain by using the knowledge of multiple seen source domains. Mainstream DG methods typically assume that the domain label of each source sample is known a priori, which is challenged to be satisfied in many real-world applications. In this paper, we study a practical problem of compound DG, which relaxes the discrete domain assumption to the mixed source domains setting. On the other hand, current DG algorithms prioritize the focus on semantic invariance across domains (one-vs-one), while paying less attention to the holistic semantic structure (many-vs-many). Such holistic semantic structure, referred to as meta-knowledge here, is crucial for learning generalizable representations. To this end, we present Compound Domain Generalization via Meta-Knowledge Encoding (COMEN), a general approach to automatically discover and model latent domains in two steps. Firstly, we introduce Style-induced Domain-specific Normalization (SDNorm) to re-normalize the multi-modal underlying distributions, thereby dividing the mixture of source domains into latent clusters. Secondly, we harness the prototype representations, the centroids of classes, to perform relational modeling in the embedding space with two parallel and complementary modules, which explicitly encode the semantic structure for the out-of-distribution generalization. Experiments on four standard DG benchmarks reveal that COMEN exceeds the state-of-the-art performance without the need of domain supervision.

Temporal grounding in videos aims to localize one target video segment that semantically corresponds to a given query sentence. Thanks to the semantic diversity of natural language descriptions, temporal grounding allows activity grounding beyond pre-defined classes and has received increasing attention in recent years. The semantic diversity is rooted in the principle of compositionality in linguistics, where novel semantics can be systematically described by combining known words in novel ways (compositional generalization). However, current temporal grounding datasets do not specifically test for the compositional generalizability. To systematically measure the compositional generalizability of temporal grounding models, we introduce a new Compositional Temporal Grounding task and construct two new dataset splits, i.e., Charades-CG and ActivityNet-CG. Evaluating the state-of-the-art methods on our new dataset splits, we empirically find that they fail to generalize to queries with novel combinations of seen words. To tackle this challenge, we propose a variational cross-graph reasoning framework that explicitly decomposes video and language into multiple structured hierarchies and learns fine-grained semantic correspondence among them. Experiments illustrate the superior compositional generalizability of our approach.

Profiting from the advance of deep convolutional networks, current state-of-the-art video action recognition models have achieved remarkable progress. Nevertheless, most of existing models suffer from low interpretability of the predicted actions. Inspired by the observation that temporally-configured human-object interactions often serve as a key indicator of many actions, this work crafts an action reasoning framework that performs Markov Logic Network (MLN) based probabilistic logical inference. Crucially, we propose to encode an action by first-order logical rules that correspond to the temporal changes of visual relationships in videos. The main contributions of this work are two-fold: 1) Different from existing black-box models, the proposed model simultaneously implements the localization of temporal boundaries and the recognition of action categories by grounding the logical rules of MLN in videos. The weight associated with each such rule further provides an estimate of confidence. These collectively make our model more explainable and robust. 2) Instead of using hand-crafted logical rules in conventional MLN, we develop a data-driven instantiation of the MLN. In specific, a hybrid learning scheme is proposed. It combines MLN's weight learning and reinforcement learning, using the former's results as a self-critic for guiding the latter's training. Additionally, by treating actions as logical predicates, the proposed framework can also be integrated with deep models for further performance boost. Comprehensive experiments on two complex video action datasets (Charades & CAD-120) clearly demonstrate the effectiveness and explainability of our proposed method.