CVPR 2022 Papers

Few-shot classification is a challenging problem that aims to learn a model that can adapt to unseen classes given a few labeled samples. Recent approaches pre-train a feature extractor, and then fine-tune for episodic meta-learning. Other methods leverage spatial features to learn pixel-level correspondence while jointly training a classifier. However, results using such approaches show marginal improvements. In this paper, inspired by the transformer style self-attention mechanism, we propose a strategy to cross-attend and re-weight discriminative features for few-shot classification. Given a base representation of support and query images after global pooling, we introduce a single shared module that projects features and cross-attends in two aspects: (i) query to support, and (ii) support to query. The module computes attention scores between features to produce an attention pooled representation of features in the same class that is later added to the original representation followed by a projection head. This effectively re-weights features in both aspects (i & ii) to produce features that better facilitate improved metric-based meta-learning. Extensive experiments on public benchmarks show our approach outperforms state-of-the-art methods by 3% 5%.

While class activation map (CAM) generated by image classification network has been widely used for weakly supervised object localization (WSOL) and semantic segmentation (WSSS), such classifiers usually focus on discriminative object regions. In this paper, we propose Contrastive learning for Class-agnostic Activation Map (C^2AM) generation only using unlabeled image data, without the involvement of image-level supervision. The core idea comes from the observation that i) semantic information of foreground objects usually differs from their backgrounds; ii) foreground objects with similar appearance or background with similar color/texture have similar representations in the feature space. We form the positive and negative pairs based on the above relations and force the network to disentangle foreground and background with a class-agnostic activation map using a novel contrastive loss. As the network is guided to discriminate cross-image foreground-background, the class-agnostic activation maps learned by our approach generate more complete object regions. We successfully extracted from C^2AM class-agnostic object bounding boxes for object localization and background cues to refine CAM generated by classification network for semantic segmentation. Extensive experiments on CUB-200-2011, ImageNet-1K, and PASCAL VOC2012 datasets show that both WSOL and WSSS can benefit from the proposed C^2AM.

Long-tail object detection suffers from poor performance on tail categories. We reveal that the real culprit lies in the extremely imbalanced distribution of the classifier's weight norm. For conventional softmax cross-entropy loss, such imbalanced weight norm distribution yields ill conditioned decision boundary for categories which have small weight norms. To get rid of this situation, we choose to maximize the cosine similarity between the learned feature and the weight vector of target category rather than the inner-product of them. The decision boundary between any two categories is the angular bisector of their weight vectors. Whereas, the absolutely equal decision boundary is suboptimal because it reduces the model's sensitivity to various categories. Intuitively, categories with rich data diversity should occupy a larger area in the classification space while categories with limited data diversity should occupy a slightly small space. Hence, we devise a Category-Aware Angular Margin Loss (C2AM Loss) to introduce an adaptive angular margin between any two categories. Specifically, the margin between two categories is proportional to the ratio of their classifiers' weight norms. As a result, the decision boundary is slightly pushed towards the category which has a smaller weight norm. We conduct comprehensive experiments on LVIS dataset. C2AM Loss brings 4.9 5.2 AP improvements on different detectors and backbones compared with baseline.

Recently, many excellent weakly supervised semantic segmentation (WSSS) works are proposed based on class activation mapping (CAM). However, there are few works that consider the characteristics of medical images. In this paper, we find that there are mainly two challenges of medical images in WSSS: i) the boundary of object foreground and background is not clear; ii) the co-occurrence phenomenon is very severe in training stage. We thus propose a Causal CAM (C-CAM) method to overcome the above challenges. Our method is motivated by two cause-effect chains including category-causality chain and anatomy-causality chain. The category-causality chain represents the image content (cause) affects the category (effect). The anatomy-causality chain represents the anatomical structure (cause) affects the organ segmentation (effect). Extensive experiments were conducted on three public medical image data sets. Our C-CAM generates the best pseudo masks with the DSC of 77.26%, 80.34% and 78.15% on ProMRI, ACDC and CHAOS compared with other CAM-like methods. The pseudo masks of C-CAM are further used to improve the segmentation performance for organ segmentation tasks. Our C-CAM achieves DSC of 83.83% on ProMRI and DSC of 87.54% on ACDC, which outperforms state-of-the-art WSSS methods. Our code is available at https://github.com/Tian-lab/C-CAM.

Modern handheld devices can acquire burst image sequence in a quick succession. However, the individual acquired frames suffer from multiple degradations and are misaligned due to camera shake and object motions. The goal of Burst Image Restoration is to effectively combine complimentary cues across multiple burst frames to generate high-quality outputs. Towards this goal, we develop a novel approach by solely focusing on the effective information exchange between burst frames, such that the degradations get filtered out while the actual scene details are preserved and enhanced. Our central idea is to create a set of pseudo-burst features that combine complimentary information from all the input burst frames to seamlessly exchange information. The pseudo-burst representations encode channel-wise features from the original burst images, thus making it easier for the model to learn distinctive information offered by multiple burst frames. However, the pseudo-burst cannot be successfully created unless the individual burst frames are properly aligned to discount inter-frame movements. Therefore, our approach initially extracts preprocessed features from each burst frame and matches them using an edge-boosting burst alignment module. The pseudo-burst features are then created and enriched using multi-scale contextual information. Our final step is to adaptively aggregate information from the pseudo-burst features to progressively increase resolution in multiple stages while merging the pseudo-burst features. In comparison to existing works that usually follow a late fusion scheme with single-stage upsampling, our approach performs favorably, delivering state-of-the-art performance on burst super-resolution, burst low-light image enhancement and burst denoising tasks. Our codes will be publicly released.

We present a learning-based framework, recurrent transformer network (RTN), to restore heavily degraded old films. Instead of performing frame-wise restoration, our method is based on the hidden knowledge learned from adjacent frames that contain abundant information about the occlusion, which is beneficial to restore challenging artifacts of each frame while ensuring temporal coherency. Moreover, contrasting the representation of the current frame and the hidden knowledge makes it possible to infer the scratch position in an unsupervised manner, and such defect localization generalizes well to real-world degradations. To better resolve mixed degradation and compensate for the flow estimation error during frame alignment, we propose to leverage more expressive transformer blocks for spatial restoration. Experiments on both synthetic dataset and real-world old films demonstrate the significant superiority of the proposed RTN over existing solutions. In addition, the same framework can effectively propagate the color from keyframes to the whole video, ultimately yielding compelling restored films.

In Class Incremental Learning (CIL), a classification model is progressively trained at each incremental step on an evolving dataset of new classes, while at the same time, it is required to preserve knowledge of all the classes observed so far. Prototypical representations can be leveraged to model feature distribution for the past data and inject information of former classes in later incremental steps without resorting to stored exemplars. However, if not updated, those representations become increasingly outdated as the incremental learning progresses with new classes. To address the aforementioned problems, we propose a framework which aims to (i) model the semantic drift by learning the relationship between representations of past and novel classes among incremental steps, and (ii) estimate the feature drift, defined as the evolution of the representations learned by models at each incremental step. Semantic and feature drifts are then jointly exploited to infer up-to-date representations of past classes (evanescent representations), and thereby infuse past knowledge into incremental training. We experimentally evaluate our framework achieving exemplar-free SotA results on multiple benchmarks. In the ablation study, we investigate nontrivial relationships between evanescent representations and models.

Most existing works in vision-and-language navigation (VLN) focus on either discrete or continuous environments, training agents that cannot generalize across the two. Although learning to navigate in continuous spaces is closer to the real-world, training such an agent is significantly more difficult than training an agent in discrete spaces. However, recent advances in discrete VLN are challenging to translate to continuous VLN due to the domain gap. The fundamental difference between the two setups is that discrete navigation assumes prior knowledge of the connectivity graph of the environment, so that the agent can effectively transfer the problem of navigation with low-level controls to jumping from node to node with high-level actions by grounding to an image of a navigable direction. To bridge the discrete-to-continuous gap, we propose a predictor to generate a set of candidate waypoints during navigation, so that agents designed with high-level actions can be transferred to and trained in continuous environments. We refine the connectivity graph of Matterport3D to fit the continuous Habitat-Matterport3D, and train the waypoints predictor with the refined graphs to produce accessible waypoints at each time step. Moreover, we demonstrate that the predicted waypoints can be augmented during training to diversify the views and paths, and therefore enhance agent's generalization ability. Through extensive experiments we show that agents navigating in continuous environments with predicted waypoints perform significantly better than agents using low-level actions, which reduces the absolute discrete-to-continuous gap by 11.76% Success Weighted by Path Length (SPL) for the Cross-Modal Matching Agent and 18.24% SPL for the Recurrent VLN-BERT. Our agents, trained with a simple imitation learning objective, outperform previous methods by a large margin, achieving new state-of-the-art results on the testing environments of the R2R-CE and the RxR-CE datasets.

Weakly supervised object localization aims to find a target object region in a given image with only weak supervision, such as image-level labels. Most existing methods use a class activation map (CAM) to generate a localization map; however, a CAM identifies only the most discriminative parts of a target object rather than the entire object region. In this work, we find the gap between classification and localization in terms of the misalignment of the directions between an input feature and a class-specific weight. We demonstrate that the misalignment suppresses the activation of CAM in areas that are less discriminative but belong to the target object. To bridge the gap, we propose a method to align feature directions with a class-specific weight. The proposed method achieves a state-of-the-art localization performance on the CUB-200-2011 and ImageNet-1K benchmarks.

Pre-training a model to learn transferable video-text representation for retrieval has attracted a lot of attention in recent years. Previous dominant works mainly adopt two separate encoders for efficient retrieval, but ignore local associations between videos and texts. Another line of research uses a joint encoder to interact video with texts, but results in low efficiency since each text-video pair needs to be fed into the model. In this work, we enable fine-grained video-text interactions while maintaining high efficiency for retrieval via a novel pretext task, dubbed as Multiple Choice Questions (MCQ), where a parametric module BridgeFormer is trained to answer the "questions" constructed by the text features via resorting to the video features. Specifically, we exploit the rich semantics of text (i.e., nouns and verbs) to build questions, with which the video encoder can be trained to capture more regional content and temporal dynamics. In the form of questions and answers, the semantic associations between local video-text features can be properly established. BridgeFormer is able to be removed for downstream retrieval, rendering an efficient and flexible model with only two encoders. Our method outperforms state-of-the-art methods on the popular text-to-video retrieval task in five datasets with different experimental setups (i.e., zero-shot and fine-tune), including HowTo100M (one million videos). We further conduct zero-shot action recognition, which can be cast as video-to-text retrieval, and our approach also significantly surpasses its counterparts. As an additional benefit, our method achieves competitive results with much shorter pre-training videos on single-modality downstream tasks, e.g., action recognition with linear evaluation.

Bridging global context interactions correctly is important for high-fidelity image completion with large masks. Previous methods attempting this via deep or large receptive field (RF) convolutions cannot escape from the dominance of nearby interactions, which may be inferior. In this paper, we propose to treat image completion as a directionless sequence-to-sequence prediction task, and deploy a transformer to directly capture long-range dependence. Crucially, we employ a restrictive CNN with small and non-overlapping RF for weighted token representation, which allows the transformer to explicitly model the long-range visible context relations with equal importance in all layers, without implicitly confounding neighboring tokens when larger RFs are used. To improve appearance consistency between visible and generated regions, a novel attention-aware layer (AAL) is introduced to better exploit distantly related high-frequency features. Overall, extensive experiments demonstrate superior performance compared to state-of-the-art methods on several datasets. Code is available at https://github.com/lyndonzheng/TFill.

3D object detection is a crucial research topic in computer vision, which usually uses 3D point clouds as input in conventional setups. Recently, there is a trend of leveraging multiple sources of input data, such as complementing the 3D point cloud with 2D images that often have richer color and fewer noises. However, due to the heterogeneous geometrics of the 2D and 3D representations, it prevents us from applying off-the-shelf neural networks to achieve multimodal fusion. To that end, we propose Bridged Transformer (BrT), an end-to-end architecture for 3D object detection. BrT is simple and effective, which learns to identify 3D and 2D object bounding boxes from both points and image patches. A key element of BrT lies in the utilization of object queries for bridging 3D and 2D spaces, which unifies different sources of data representations in Transformer. We adopt a form of feature aggregation realized by point-to-patch projections which further strengthen the interaction between images and points. Moreover, BrT works seamlessly for fusing the point cloud with multi-view images. We experimentally show that BrT surpasses state-of-the-art methods on SUN RGB-D and ScanNetV2 datasets.

Action recognition models have shown a promising capability to classify human actions in short video clips. In a real scenario, multiple correlated human actions commonly occur in particular orders, forming semantically meaningful human activities. Conventional action recognition approaches focus on analyzing single actions. However, they fail to fully reason about the contextual relations between adjacent actions, which provide potential temporal logic for understanding long videos. In this paper, we propose a prompt-based framework, Bridge-Prompt (Br-Prompt), to model the semantics across adjacent actions, so that it simultaneously exploits both out-of-context and contextual information from a series of ordinal actions in instructional videos. More specifically, we reformulate the individual action labels as integrated text prompts for supervision, which bridge the gap between individual action semantics. The generated text prompts are paired with corresponding video clips, and together co-train the text encoder and the video encoder via a contrastive approach. The learned vision encoder has a stronger capability for ordinal-action-related downstream tasks, e.g. action segmentation and human activity recognition. We evaluate the performances of our approach on several video datasets: Georgia Tech Egocentric Activities (GTEA), 50Salads, and the Breakfast dataset. Br-Prompt achieves state-of-the-art on multiple benchmarks. Code is available at: https://github.com/ttlmh/Bridge-Prompt.

Despite recent advances in deep neural models for semantic image editing, present approaches are dependent on explicit human input. Previous work assumes the availability of manually curated datasets for supervised learning, while for unsupervised approaches the human inspection of discovered components is required to identify those which modify worthwhile semantic features. Here, we present a novel alternative: the utilization of brain responses as a supervision signal for learning semantic feature representations. Participants (N=30) in a neurophysiological experiment were shown artificially generated faces and instructed to look for a particular semantic feature, such as "old" or "smiling", while their brain responses were recorded via electroencephalography (EEG). Using supervision signals inferred from these responses, semantic features within the latent space of a generative adversarial network (GAN) were learned and then used to edit semantic features of new images. We show that implicit brain supervision achieves comparable semantic image editing performance to explicit manual labeling. This work demonstrates the feasibility of utilizing implicit human reactions recorded via brain-computer interfaces for semantic image editing and interpretation.

Recently, Multilayer Perceptron (MLP) becomes the hotspot in the field of computer vision tasks. Without inductive bias, MLPs perform well on feature extraction and achieve amazing results. However, due to the simplicity of their structures, the performance highly depends on the local features communication machenism. To further improve the performance of MLP, we introduce information communication mechanisms from brain-inspired neural networks. Spiking Neural Network (SNN) is the most famous brain-inspired neural network, and achieve great success on dealing with sparse data. Leaky Integrate and Fire (LIF) neurons in SNNs are used to communicate between different time steps. In this paper, we incorporate the machanism of LIF neurons into the MLP models, to achieve better accuracy without extra FLOPs. We propose a full-precision LIF operation to communicate between patches, including horizontal LIF and vertical LIF in different directions. We also propose to use group LIF to extract better local features. With LIF modules, our SNN-MLP model achieves 81.9%, 83.3% and 83.5% top-1 accuracy on ImageNet dataset with only 4.4G, 8.5G and 15.2G FLOPs, respectively, which are state-of-the-art results as far as we know.

Deep neural networks are vulnerable to Trojan attacks. Existing attacks use visible patterns (e.g., a patch or image transformations) as triggers, which are vulnerable to human inspection. In this paper, we propose stealthy and efficient Trojan attacks, BppAttack. Based on existing biology literature on human visual systems, we propose to use image quantization and dithering as the Trojan trigger, making imperceptible changes. It is a stealthy and efficient attack without training auxiliary models. Due to the small changes made to images, it is hard to inject such triggers during training. To alleviate this problem, we propose a contrastive learning based approach that leverages adversarial attacks to generate negative sample pairs so that the learned trigger is precise and accurate. The proposed method achieves high attack success rates on four benchmark datasets, including MNIST, CIFAR-10, GTSRB, and CelebA. It also effectively bypasses existing Trojan defenses and human inspection. Our code can be found in https://github.com/RU-System-Software-and-Security/BppAttack.

In this paper, we propose a simple attention mechanism, we call Box-Attention. It enables spatial interaction between grid features, as sampled from boxes of interest, and improves the learning capability of transformers for several vision tasks. Specifically, we present BoxeR, short for Box Transformer, which attends to a set of boxes by predicting their transformation from a reference window on an input feature map. The BoxeR computes attention weights on these boxes by considering its grid structure. Notably, BoxeR-2D naturally reasons about box information within its attention module, making it suitable for end-to-end instance detection and segmentation tasks. By learning invariance to rotation in the box-attention module, BoxeR-3D is capable of generating discriminative information from a bird's-eye view plane for 3D end-to-end object detection. Our experiments demonstrate that the proposed BoxeR-2D achieves state-of-the-art results on COCO detection and instance segmentation. Besides, BoxeR-3D improves over the end-to-end 3D object detection baseline and already obtains a compelling performance for the vehicle category of Waymo Open, without any class-specific optimization. Code is available at https://github.com/kienduynguyen/BoxeR.

We propose a novel adversarial attack targeting content features in some deep layer, that is, individual neurons in the layer. A naive method that enforces a fixed value/percentage bound for neuron activation values can hardly work and generates very noisy samples. The reason is that the level of perceptual variation entailed by a fixed value bound is non-uniform across neurons and even for the same neuron. We hence propose a novel distribution quantile bound for activation values and a polynomial barrier loss function. Given a benign input, a fixed quantile bound is translated to many value bounds, one for each neuron, based on the distributions of the neuron's activations and the current activation value on the given input. These individualized bounds enable fine-grained regulation, allowing content feature mutations with bounded perceptional variations. Our evaluation on ImageNet and five different model architectures demonstrates that our attack is effective. Compared to seven other latest adversarial attacks in both the pixel space and the feature space, our attack can achieve the state-of-the-art trade-off between attack success rate and imperceptibility.