CVPR 2022 Papers

Recent studies have shown that adversarial examples hand-crafted on one white box model can be used to attack other black-box models. Such cross-model transferability makes it feasible to perform black-box attacks, which has raised security concerns for real-world DNNs applications. Nevertheless, existing works mostly focus on investigating the adversarial transferability across different deep models that share the same modality of input data. The cross-modal transferability of adversarial perturbation has never been explored. This paper investigates the transferability of adversarial perturbation across different modalities, i.e., leveraging adversarial perturbation generated on white-box image models to attack black-box video models. Specifically, motivated by the observation that the low-level feature space between images and video frames are similar, we propose a simple yet effective cross-modal attack method, named as Image To Video (I2V) attack. I2V generates adversarial frames by minimizing the cosine similarity between features of pre-trained image models from adversarial and benign examples, then combines the generated adversarial frames to perform black-box attacks on video recognition models. Extensive experiments demonstrate that I2V can achieve high attack success rates on different black-box video recognition models. On Kinetics-400 and UCF-101, I2V achieves an average attack success rate of 77.88% and 65.68%, respectively, which sheds light on the feasibility of cross-modal adversarial attacks.

We present a cross-modal Transformer-based framework, which jointly encodes video data and text labels for zero-shot action recognition (ZSAR). Our model employs a conceptually new pipeline by which visual representations are learned in conjunction with visual-semantic associations in an end-to-end manner. The model design provides a natural mechanism for visual and semantic representations to be learned in a shared knowledge space, whereby it encourages the learned visual embedding to be discriminative and more semantically consistent. In zero-shot inference, we devise a simple semantic transfer scheme that embeds semantic relatedness information between seen and unseen classes to composite unseen visual prototypes. Accordingly, the discriminative features in the visual structure could be preserved and exploited to alleviate the typical zero-shot issues of information loss, semantic gap, and the hubness problem. Under a rigorous zero-shot setting of not pre-training on additional datasets, the experiment results show our model considerably improves upon the state of the arts in ZSAR, reaching encouraging top-1 accuracy on UCF101, HMDB51, and ActivityNet benchmark datasets. Code will be made available.

This work digs into a root question in human perception: can face geometry be gleaned from one's voices? Previous works that study this question only adopt developments in image synthesis and convert voices into face images to show correlations, but working on the image domain unavoidably involves predicting attributes that voices cannot hint, including facial textures, hairstyles, and backgrounds. We instead investigate the ability to reconstruct 3D faces to concentrate on only geometry, which is much more physiologically grounded. We propose our analysis framework, Cross-Modal Perceptionist, under both supervised and unsupervised learning. First, we construct a dataset, Voxceleb-3D, which extends Voxceleb and includes paired voices and face meshes, making supervised learning possible. Second, we use a knowledge distillation mechanism to study whether face geometry can still be gleaned from voices without paired voices and 3D face data under limited availability of 3D face scans. We break down the core question into four parts and perform visual and numerical analyses as responses to the core question. Our findings echo those in physiology and neuroscience about the correlation between voices and facial structures. The work provides future human-centric cross-modal learning with explainable foundations. See our project page.

We consider the problem of Vision-and-Language Navigation (VLN). The majority of current methods for VLN are trained end-to-end using either unstructured memory such as LSTM, or using cross-modal attention over the egocentric observations of the agent. In contrast to other works, our key insight is that the association between language and vision is stronger when it occurs in explicit spatial representations. In this work, we propose a cross-modal map learning model for vision-and-language navigation that first learns to predict the top-down semantics on an egocentric map for both observed and unobserved regions, and then predicts a path towards the goal as a set of waypoints. In both cases, the prediction is informed by the language through cross-modal attention mechanisms. We experimentally test the basic hypothesis that language-driven navigation can be solved given a map, and then show competitive results on the full VLN-CE benchmark.

Automatic generation of ophthalmic reports using data-driven neural networks has great potential in clinical practice. When writing a report, ophthalmologists make inferences with prior clinical knowledge. This knowledge has been neglected in prior medical report generation methods. To endow models with the capability of incorporating expert knowledge, we propose a Cross-modal clinical Graph Transformer (CGT) for ophthalmic report generation (ORG), in which clinical relation triples are injected into the visual features as prior knowledge to drive the decoding procedure. However, two major common Knowledge Noise (KN) issues may affect models' effectiveness. 1) Existing general biomedical knowledge bases such as the UMLS may not align meaningfully to the specific context and language of the report, limiting their utility for knowledge injection. 2) Incorporating too much knowledge may divert the visual features from their correct meaning. To overcome these limitations, we design an automatic information extraction scheme based on natural language processing to obtain clinical entities and relations directly from in-domain training reports. Given a set of ophthalmic images, our CGT first restores a sub-graph from the clinical graph and injects the restored triples into visual features. Then visible matrix is employed during the encoding procedure to limit the impact of knowledge. Finally, reports are predicted by the encoded cross-modal features via a Transformer decoder. Extensive experiments on the large-scale FFA-IR benchmark demonstrate that the proposed CGT is able to outperform previous benchmark methods and achieve state-of-the-art performances.

Audiovisual Event (AVE) localization requires the model to jointly localize an event by observing audio and visual information. However, in unconstrained videos, both information types may be inconsistent or suffer from severe background noise. Hence this paper proposes a novel cross-modal background suppression network for AVE task, operating at the time- and event-level, aiming to improve localization performance through suppressing asynchronous audiovisual background frames from the examined events and reducing redundant noise. Specifically, the time-level background suppression scheme forces the audio and visual modality to focus on the related information in the temporal dimension that the opposite modality considers essential, and reduces attention to the segments that the other modal considers as background. The event-level background suppression scheme uses the class activation sequences predicted by audio and visual modalities to control the final event category prediction, which can effectively suppress noise events occurring accidentally in a single modality. Furthermore, we introduce a cross-modal gated attention scheme to extract relevant visual regions from complex scenes exploiting both global visual and audio signals. Extensive experiments show our method outperforms the state-of-the-art methods by a large margin in both supervised and weakly supervised AVE settings.

Current Knowledge Distillation (KD) methods for semantic segmentation often guide the student to mimic the teacher's structured information generated from individual data samples. However, they ignore the global semantic relations among pixels across various images that are valuable for KD. This paper proposes a novel Cross-Image Relational KD (CIRKD), which focuses on transferring structured pixel-to-pixel and pixel-to-region relations among the whole images. The motivation is that a good teacher network could construct a well-structured feature space in terms of global pixel dependencies. CIRKD makes the student mimic better structured semantic relations from the teacher, thus improving the segmentation performance. Experimental results over Cityscapes, CamVid and Pascal VOC datasets demonstrate the effectiveness of our proposed approach against state-of-the-art distillation methods. The code is available at https://github.com/winycg/CIRKD.

In this paper, we look at the problem of cross-domain few-shot classification that aims to learn a classifier from previously unseen classes and domains with few labeled samples. Recent approaches broadly solve this problem by parameterizing their few-shot classifiers with task-agnostic and task-specific weights where the former is typically learned on a large training set and the latter is dynamically predicted through an auxiliary network conditioned on a small support set. In this work, we focus on the estimation of the latter, and propose to learn task-specific weights from scratch directly on a small support set, in contrast to dynamically estimating them. In particular, through systematic analysis, we show that task-specific weights through parametric adapters in matrix form with residual connections to multiple intermediate layers of a backbone network significantly improves the performance of the state-of-the-art models in the Meta-Dataset benchmark with minor additional cost.

The performance of nighttime semantic segmentation is restricted by the poor illumination and a lack of pixel-wise annotation, which severely limit its application in autonomous driving. Existing works, e.g., using the twilight as the intermediate target domain to perform the adaptation from daytime to nighttime, may fail to cope with the inherent difference between datasets caused by the camera equipment and the urban style. Faced with these two types of domain shifts, i.e., the illumination and the inherent difference of the datasets, we propose a novel domain adaptation framework via cross-domain correlation distillation, called CCDistill. The invariance of illumination or inherent difference between two images is fully explored so as to make up for the lack of labels for nighttime images. Specifically, we extract the content and style knowledge contained in features, calculate the degree of inherent or illumination difference between two images. The domain adaptation is achieved using the invariance of the same kind of difference. Extensive experiments on Dark Zurich and ACDC demonstrate that CCDistill achieves the state-of-the-art performance for nighttime semantic segmentation. Notably, our method is a one-stage domain adaptation network which can avoid affecting the inference time. Our implementation is available at https://github.com/ghuan99/CCDistill.

We address the task of domain adaptation in object detection, where there is a domain gap between a domain with annotations (source) and a domain of interest without annotations (target). As an effective semi-supervised learning method, the teacher-student framework (a student model is supervised by the pseudo labels from a teacher model) has also yielded a large accuracy gain in cross-domain object detection. However, it suffers from the domain shift and generates many low-quality pseudo labels (e.g., false positives), which leads to sub-optimal performance. To mitigate this problem, we propose a teacher-student framework named Adaptive Teacher (AT) which leverages domain adversarial learning and weak-strong data augmentation to address the domain gap. Specifically, we employ feature-level adversarial training in the student model, allowing features derived from the source and target domains to share similar distributions. This process ensures the student model produces domain-invariant features. Furthermore, we apply weak-strong augmentation and mutual learning between the teacher model (taking data from the target domain) and the student model (taking data from both domains). This enables the teacher model to learn the knowledge from the student model without being biased to the source domain. We show that AT demonstrates superiority over existing approaches and even Oracle (fully-supervised) models by a large margin. For example, we achieve 50.9% (49.3%) mAP on Foggy Cityscape (Clipart1K), which is 9.2% (5.2%) and 8.2% (11.0%) higher than previous state-of-the-art and Oracle, respectively.

In this paper, we present a new cross-architecture contrastive learning (CACL) framework for self-supervised video representation learning. CACL consists of a 3D CNN and a video transformer which are used in parallel to generate diverse positive pairs for contrastive learning. This allows the model to learn strong representations from such diverse yet meaningful pairs. Furthermore, we introduce a temporal self-supervised learning module able to predict an Edit distance explicitly between two video sequences in the temporal order. This enables the model to learn a rich temporal representation that compensates strongly to the video-level representation learned by the CACL. We evaluate our method on the tasks of video retrieval and action recognition on UCF101 and HMDB51 datasets, where our method achieves excellent performance, surpassing the state-of-the-art methods such as VideoMoCo and MoCo+BE by a large margin.

Profiting from large-scale training datasets, advances in neural architecture design and efficient inference, joint embeddings have become the dominant approach for tackling cross-modal retrieval. In this work we first show that, despite their effectiveness, state-of-the-art joint embeddings suffer significantly from the longstanding "hubness problem" in which a small number of gallery embeddings form the nearest neighbours of many queries. Drawing inspiration from the NLP literature, we formulate a simple but effective framework called Querybank Normalisation (QB-Norm) that re-normalises query similarities to account for hubs in the embedding space. QB-Norm improves retrieval performance without requiring retraining. Differently from prior work, we show that QB-Norm works effectively without concurrent access to any test set queries. Within the QB-Norm framework, we also propose a novel similarity normalisation method, the Dynamic Inverted Softmax, that is significantly more robust than existing approaches. We showcase QB-Norm across a range of cross modal retrieval models and benchmarks where it consistently enhances strong baselines beyond the state of the art. Code is available at https://vladbogo.github.io/QB-Norm/.

Cross domain object detection is a realistic and challenging task in the wild. It suffers from performance degradation due to large shift of data distributions and lack of instance-level annotations in the target domain. Existing approaches mainly focus on either of these two difficulties, even though they are closely coupled in cross domain object detection. To solve this problem, we propose a novel Target-perceived Dual-branch Distillation (TDD) framework. By integrating detection branches of both source and target domains in a unified teacher-student learning scheme, it can reduce domain shift and generate reliable supervision effectively. In particular, we first introduce a distinct Target Proposal Perceiver between two domains. It can adaptively enhance source detector to perceive objects in a target image, by leveraging target proposal contexts from iterative cross-attention. Afterwards, we design a concise Dual Branch Self Distillation strategy for model training, which can progressively integrate complementary object knowledge from different domains via self-distillation in two branches. Finally, we conduct extensive experiments on a number of widely-used scenarios in cross domain object detection. The results show that our TDD significantly outperforms the state-of-the-art methods on all the benchmarks. The codes and models will be released afterwards.

Learning-based depth estimation has witnessed recent progress in multiple directions; from self-supervision using monocular video to supervised methods offering highest accuracy. Complementary to supervision, further boosts to performance and robustness are gained by combining information from multiple signals. In this paper we systematically investigate key trade-offs associated with sensor and modality design choices as well as related model training strategies. Our study leads us to a new method, capable of connecting modality-specific advantages from polarisation, Time-of-Flight and structured-light inputs. We propose a novel pipeline capable of estimating depth from monocular polarisation for which we evaluate various training signals. The inversion of differentiable analytic models thereby connects scene geometry with polarisation and ToF signals and enables self-supervised and cross-modal learning. In the absence of existing multimodal datasets, we examine our approach with a custom-made multi-modal camera rig and collect CroMo; the first dataset to consist of synchronized stereo polarisation, indirect ToF and structured-light depth, captured at video rates. Extensive experiments on challenging video scenes confirm both qualitative and quantitative pipeline advantages where we are able to outperform competitive monocular depth estimation methods.

Neural implicit functions have recently shown promising results on surface reconstructions from multiple views. However, current methods still suffer from excessive time complexity and poor robustness when reconstructing unbounded or complex scenes. In this paper, we present RegSDF, which shows that proper point cloud supervisions and geometry regularizations are sufficient to produce high-quality and robust reconstruction results. Specifically, RegSDF takes an additional oriented point cloud as input, and optimizes a signed distance field and a surface light field within a differentiable rendering framework. We also introduce the two critical regularizations for this optimization. The first one is the Hessian regularization that smoothly diffuses the signed distance values to the entire distance field given noisy and incomplete input. And the second one is the minimal surface regularization that compactly interpolates and extrapolates the missing geometry. Extensive experiments are conducted on DTU, BlendedMVS, and Tanks and Temples datasets. Compared with recent neural surface reconstruction approaches, RegSDF is able to reconstruct surfaces with fine details even for open scenes with complex topologies and unstructured camera trajectories.

We exploit the complementary strengths of vision and proprioception to develop a point-goal navigation system for legged robots, called VP-Nav. Legged systems are capable of traversing more complex terrain than wheeled robots, but to fully utilize this capability, we need a high-level path planner in the navigation system to be aware of the walking capabilities of the low-level locomotion policy in varying environments. We achieve this by using proprioceptive feedback to ensure the safety of the planned path by sensing unexpected obstacles like glass walls, terrain properties like slipperiness or softness of the ground and robot properties like extra payload that are likely missed by vision. The navigation system uses onboard cameras to generate an occupancy map and a corresponding cost map to reach the goal. A fast marching planner then generates a target path. A velocity command generator takes this as input to generate the desired velocity for the walking policy. A safety advisor module adds sensed unexpected obstacles to the occupancy map and environment-determined speed limits to the velocity command generator. We show superior performance compared to wheeled robot baselines, and ablation studies which have disjoint high-level planning and low-level control. We also show the real-world deployment of VP-Nav on a quadruped robot with onboard sensors and computation. Videos at https://navigation-locomotion.github.io

We address the task of 6D multi-object pose: given a set of known 3D objects and an RGB or RGB-D input image, we detect and estimate the 6D pose of each object. We propose a new approach to 6D object pose estimation which consists of an end-to-end differentiable architecture that makes use of geometric knowledge. Our approach iteratively refines both pose and correspondence in a tightly coupled manner, allowing us to dynamically remove outliers to improve accuracy. We use a novel differentiable layer to perform pose refinement by solving an optimization problem we refer to as Bidirectional Depth-Augmented Perspective-N-Point (BD-PnP). Our method achieves state-of-the-art accuracy on standard 6D Object Pose benchmarks.

Since the rise of vision-language navigation (VLN), great progress has been made in instruction following -- building a follower to navigate environments under the guidance of instructions. However, far less attention has been paid to the inverse task: instruction generation -- learning a speaker to generate grounded descriptions for navigation routes. Existing VLN methods train a speaker independently and typically treat it as a data augmentation tool for strengthening the follower, while ignoring rich cross-task relations. Here we describe an approach that learns the two tasks simultaneously and exploits their intrinsic correlations to boost the training of each: the follower judges whether the speaker-created instruction explains the original navigation route correctly, and vice versa. Without the need of aligned instruction-path pairs, such cycle-consistent learning scheme is complementary to task-specific training objectives defined on labeled data, and can also be applied over unlabeled paths (sampled without paired instructions). Another agent, called creator, is added to generate counterfactual environments. It greatly changes current scenes yet leaves novel items -- which are crucial for the execution of original instructions -- unchanged. Thus more informative training scenes are synthesized and the three agents compose a powerful VLN learning system. Experiments on a standard benchmark show that our approach improves the performance of various follower models and produces accurate navigation instructions. Our code will be released.

Robustness and discrimination power are two fundamental requirements in visual object tracking. In most tracking paradigms, we find that the features extracted by the popular Siamese-like networks can not fully discriminatively model the tracked targets and distractor objects, hindering them from simultaneously meeting these two requirements. While most methods focus on designing robust matching operations, we propose a novel target-dependent feature network inspired by the self-/cross-attention scheme. In contrast to the Siamese-like feature extraction, our network deeply embeds cross-image feature correlation in multiple layers of the feature network. By extensively matching the features of the two images through multiple layers, it is able to suppress non-target features, resulting in instance-varying feature extraction. The output features of the search image can be directly used for predicting target locations without extra correlation step. Moreover, our model can be flexibly pre-trained on abundant unpaired images, leading to notably faster convergence than the existing methods. Extensive experiments show our method achieves the state-of-the-art results while running at real-time. Our feature networks also can be applied to existing tracking pipelines seamlessly to raise the tracking performance.