ICCV 2023 论文列表

Vision Transformers (ViTs) have demonstrated powerful representation ability in various visual tasks thanks to their intrinsic data-hungry nature. However, we unexpectedly find that ViTs perform vulnerably when applied to face recognition (FR) scenarios with extremely large datasets. We investigate the reasons for this phenomenon and discover that the existing data augmentation approach and hard sample mining strategy are incompatible with ViTs-based FR backbone due to the lack of tailored consideration on preserving face structural information and leveraging each local token information. To remedy these problems, this paper proposes a superior FR model called TransFace, which employs a patch-level data augmentation strategy named DPAP and a hard sample mining strategy named EHSM. Specially, DPAP randomly perturbs the amplitude information of dominant patches to expand sample diversity, which effectively alleviates the overfitting problem in ViTs. EHSM utilizes the information entropy in the local tokens to dynamically adjust the importance weight of easy and hard samples during training, leading to a more stable prediction. Experiments on several benchmarks demonstrate the superiority of our TransFace. Code and models are available at https://github.com/DanJun6737/TransFace.

3D multi-object tracking (MOT) is vital for many applications including autonomous driving vehicles and service robots. With the commonly used tracking-by-detection paradigm, 3D MOT has made important progress in recent years. However, these methods only use the detection boxes of the current frame to obtain trajectory-box association results, which makes it impossible for the tracker to recover objects missed by the detector. In this paper, we present TrajectoryFormer, a novel point-cloud-based 3D MOT framework. To recover the missed object by detector, we generates multiple trajectory hypotheses with hybrid candidate boxes, including temporally predicted boxes and currentframe detection boxes, for trajectory-box association. The predicted boxes can propagate object's history trajectory information to the current frame and thus the network can tolerate short-term miss detection of the tracked objects. We combine long-term object motion feature and short-term object appearance feature to create per-hypothesis feature embedding, which reduces the computational overhead for spatial-temporal encoding. Additionally, we introduce a Global-Local Interaction Module to conduct information interaction among all hypotheses and models their spatial relations, leading to accurate estimation of hypotheses. Our TrajectoryFormer achieves state-of-the-art performance on the Waymo 3D MOT benchmarks.

Pedestrian trajectory prediction is an essentially connecting link to understanding human behavior. Recent works achieve state-of-the-art performance gained from the hand-designed post-processing, e.g., clustering. However, this post-processing suffers from expensive inference time and neglects the probability of the predicted trajectory disturbing downstream safety decisions. In this paper, we present Trajectory Unified TRansformer, called TUTR, which unifies the trajectory prediction components, social interaction and multimodal trajectory prediction, into a transformer encoder-decoder architecture to effectively remove the need for post-processing. Specifically, TUTR parses the relationships across various motion modes by an explicit global prediction and an implicit mode-level transformer encoder. Then, TUTR attends to the social interactions with neighbors by a social-level transformer decoder. Finally, a dual prediction forecasts diverse trajectories and corresponding probabilities in parallel without post-processing. TUTR achieves state-of-the-art accuracy performance and about 10x - 40x inference speed improvements compared with previous well-tuning state-of-the-art methods using post-processing.

Robust pedestrian trajectory forecasting is crucial to developing safe autonomous vehicles. Although previous works have studied adversarial robustness in the context of trajectory forecasting, some significant issues remain unaddressed. In this work, we try to tackle these crucial problems. Firstly, the previous definitions of robustness in trajectory prediction are ambiguous. We thus provide formal definitions for two kinds of robustness, namely label robustness and pure robustness. Secondly, as previous works fail to consider robustness about all points in a disturbance interval, we utilise a probably approximately correct (PAC) framework for robustness verification. Additionally, this framework can not only identify potential counterexamples, but also provides interpretable analyses of the original methods. Our approach is applied using a prototype tool named TrajPAC. With TrajPAC, we evaluate the robustness of four state-of-the-art trajectory prediction models -- Trajectron++, MemoNet, AgentFormer, and MID -- on trajectories from five scenes of the ETH/UCY dataset and scenes of the Stanford Drone Dataset. Using our framework, we also experimentally study various factors that could influence robustness performance.

Trajectory prediction has been a crucial task in building a reliable autonomous driving system by anticipating possible dangers. One key issue is to generate consistent trajectory predictions without colliding. To overcome the challenge, we propose an efficient masked autoencoder for trajectory prediction (Traj-MAE) that better represents the complicated behaviors of agents in the driving environment. Specifically, our Traj-MAE employs diverse masking strategies to pre-train the trajectory encoder and map encoder, allowing for the capture of social and temporal information among agents while leveraging the effect of environment from multiple granularities. To address the catastrophic forgetting problem that arises when pre-training the network with multiple masking strategies, we introduce a continual pre-training framework, which can help Traj-MAE learn valuable and diverse information from various strategies efficiently. Our experimental results in both multi-agent and single-agent settings demonstrate that Traj-MAE achieves competitive results with state-of-the-art methods and significantly outperforms our baseline model. Project page: https://jiazewang.com/projects/trajmae.html.

Unsupervised learning is a challenging task due to the lack of labels. Multiple Object Tracking (MOT), which inevitably suffers from mutual object interference, occlusion, etc., is even more difficult without label supervision. In this paper, we explore the latent consistency of sample features across video frames and propose an Unsupervised Contrastive Similarity Learning method, named UCSL, including three contrast modules: self-contrast, cross-contrast, and ambiguity contrast. Specifically, i) self-contrast uses intra-frame direct and inter-frame indirect contrast to obtain discriminative representations by maximizing self-similarity. ii) Cross-contrast aligns cross- and continuous-frame matching results, mitigating the persistent negative effect caused by object occlusion. And iii) ambiguity contrast matches ambiguous objects with each other to further increase the certainty of subsequent object association through an implicit manner. On existing benchmarks, our method outperforms the existing unsupervised methods using only limited help from ReID head, and even provides higher accuracy than lots of fully supervised methods.

The main challenge of Tracking by Natural Language Specification (TNL) is to predict the movement of the target object by giving two heterogeneous information, e.g., one is the static description of the main characteristics of a video contained in the textual query, i.e., long-term context; the other one is an image patch containing the object and its surroundings cropped from the current frame, i.e., the search area. Currently, most methods still struggle with the rationality of using those two information and simply fusing the two. However, the linguistic information contained in the textual query and the visual representation stored in the search area may sometimes be inconsistent, in which case the direct fusion of the two may lead to conflicts. To address this problem, we propose DecoupleTNL, introducing a video clip containing short-term context information into the framework of TNL and exploring a proper way to reduce the impact when visual representation is inconsistent with linguistic information. Concretely, we design two jointly optimized tasks, i.e., short-term context-matching and long-term context-perceiving. The context-matching task aims to gather the dynamic short-term context information in a period, while the context-perceiving task tends to extract the static long-term context information. After that, we design a long short-term modulation module to integrate both context information for accurate tracking. Extensive experiments have been conducted on three tracking benchmark datasets to demonstrate the superiority of DecoupleTNL

Most model-free visual object tracking methods formulate the tracking task as object location estimation given by a 2D segmentation or a bounding box in each video frame. We argue that this representation is limited and instead propose to guide and improve 2D tracking with an explicit object representation, namely the textured 3D shape and 6DoF pose in each video frame. Our representation tackles a complex long-term dense correspondence problem between all 3D points on the object for all video frames, including frames where some points are invisible. To achieve that, the estimation is driven by re-rendering the input video frames as well as possible through differentiable rendering, which has not been used for tracking before. The proposed optimization minimizes a novel loss function to estimate the best 3D shape, texture, and 6DoF pose. We improve the state-of-the-art in 2D segmentation tracking on three different datasets with mostly rigid objects.

Training data for video segmentation are expensive to annotate. This impedes extensions of end-to-end algorithms to new video segmentation tasks, especially in large-vocabulary settings. To 'track anything' without training on video data for every individual task, we develop a decoupled video segmentation approach (DEVA), composed of task-specific image-level segmentation and class/task-agnostic bi-directional temporal propagation. Due to this design, we only need an image-level model for the target task (which is cheaper to train) and a universal temporal propagation model which is trained once and generalizes across tasks. To effectively combine these two modules, we use bi-directional propagation for (semi-)online fusion of segmentation hypotheses from different frames to generate a coherent segmentation. We show that this decoupled formulation compares favorably to end-to-end approaches in several data-scarce tasks including large-vocabulary video panoptic segmentation, open-world video segmentation, referring video segmentation, and unsupervised video object segmentation. Code is available at: https://hkchengrex.github.io/Tracking-Anything-with-DEVA.

The field of multi-object tracking has recently seen a renewed interest in the good old schema of tracking-by-detection, as its simplicity and strong priors spare it from the complex design and painful babysitting of tracking-by-attention approaches. In view of this, we aim at extending tracking-by-detection to multi-modal settings, where a comprehensive cost has to be computed from heterogeneous information e.g., 2D motion cues, visual appearance, and pose estimates. More precisely, we follow a case study where a rough estimate of 3D information is also available and must be merged with other traditional metrics (e.g., the IoU). To achieve that, recent approaches resort to either simple rules or complex heuristics to balance the contribution of each cost. However, i) they require careful tuning of tailored hyperparameters on a hold-out set, and ii) they imply these costs to be independent, which does not hold in reality. We address these issues by building upon an elegant probabilistic formulation, which considers the cost of a candidate association as the negative log-likelihood yielded by a deep density estimator, trained to model the conditional joint probability distribution of correct associations. Our experiments, conducted on both simulated and real benchmarks, show that our approach consistently enhances the performance of several tracking-by-detection algorithms.

Deep neural networks are vulnerable to adversarial attacks. In this paper, we take the role of investigators who want to trace the attack and identify the source, that is, the particular model which the adversarial examples are generated from. Techniques derived would aid forensic investigation of attack incidents and serve as deterrence to potential attacks. We consider the buyers-seller setting where a machine learning model is to be distributed to various buyers and each buyer receives a slightly different copy with same functionality. A malicious buyer generates adversarial examples from a particular copy "Mi" and uses them to attack other copies. From these adversarial examples, the investigator wants to identify the source "Mi". To address this problem, we propose a two-stage separate-and-trace framework. The model separation stage generates multiple copies of a model for a same classification task. This process injects unique characteristics into each copy so that adversarial examples generated have distinct and traceable features. We give a parallel structure which pairs a unique tracer with the original classification model in each copy and a variational autoencoder (VAE)-based training method to achieve this goal. The tracing stage takes in adversarial examples and a few candidate models, and identifies the likely source. Based on the unique features induced by the tracer, we could effectively trace the potential adversarial copy by considering the output logits from each tracer. Empirical results show that it is possible to trace the origin of the adversarial example and the mechanism can be applied to a wide range of architectures and datasets.

Monocular depth estimation is scale-ambiguous, and thus requires scale supervision to produce metric predictions. Even so, the resulting models will be geometry-specific, with learned scales that cannot be directly transferred across domains. Because of that, recent works focus instead on relative depth, eschewing scale in favor of improved up-to-scale zero-shot transfer. In this work we introduce ZeroDepth, a novel monocular depth estimation framework capable of predicting metric scale for arbitrary test images from different domains and camera parameters. This is achieved by (i) the use of input-level geometric embeddings that enable the network to learn a scale prior over objects; and (ii) decoupling the encoder and decoder stages, via a variational latent representation that is conditioned on single frame information. We evaluated ZeroDepth targeting both outdoor (KITTI, DDAD, nuScenes) and indoor (NYUv2) benchmarks, and achieved a new state-of-the-art in both settings using the same pre-trained model, outperforming methods that train on in-domain data and require test-time scaling to produce metric estimates.

Testing the full autonomy system in simulation is the safest and most scalable way to evaluate autonomous vehicle performance before deployment. This requires simulating sensor inputs such as LiDAR. To be effective, it is essential that the simulation has low domain gap with the real world. That is, the autonomy system in simulation should perform exactly the same way it would in the real world for the same scenario. To date, there has been limited analysis into what aspects of LiDAR phenomena affect autonomy performance. It is also difficult to evaluate the domain gap of existing LiDAR simulators, as they operate on fully synthetic scenes. In this paper, we propose a novel "paired-scenario" approach to evaluating the domain gap of a LiDAR simulator by reconstructing digital twins of real world scenarios. We can then simulate LiDAR in the scene and compare it to the real LiDAR. We leverage this setting to analyze what aspects of LiDAR simulation, such as pulse phenomena, scanning effects, and asset quality, affect the domain gap with respect to the autonomy system, including perception, prediction, and motion planning, and analyze how modifications to the simulated LiDAR influence each part. We identify key aspects that are important to model, such as motion blur, material reflectance, and the accurate geometric reconstruction of traffic participants. This helps provide research directions for improving LiDAR simulation and autonomy robustness to these effects. For more information, please visit the project website: https://waabi.ai/lidar-dg

Visual recognition models are not invariant to viewpoint changes in the 3D world, as different viewing directions can dramatically affect the predictions given the same object. Although many efforts have been devoted to making neural networks invariant to 2D image translations and rotations, viewpoint invariance is rarely investigated. As most models process images in the perspective view, it is challenging to impose invariance to 3D viewpoint changes based only on 2D inputs. Motivated by the success of adversarial training in promoting model robustness, we propose Viewpoint-Invariant Adversarial Training (VIAT) to improve viewpoint robustness of common image classifiers. By regarding viewpoint transformation as an attack, VIAT is formulated as a minimax optimization problem, where the inner maximization characterizes diverse adversarial viewpoints by learning a Gaussian mixture distribution based on a new attack GMVFool, while the outer minimization trains a viewpoint-invariant classifier by minimizing the expected loss over the worst-case adversarial viewpoint distributions. To further improve the generalization performance, a distribution sharing strategy is introduced leveraging the transferability of adversarial viewpoints across objects. Experiments validate the effectiveness of VIAT in improving the viewpoint robustness of various image classifiers based on the diversity of adversarial viewpoints generated by GMVFool.

Autonomous vehicles (AV) require that neural networks used for perception be robust to different viewpoints if they are to be deployed across many types of vehicles without the repeated cost of data collection and labeling for each. AV companies typically focus on collecting data from diverse scenarios and locations, but not camera rig configurations, due to cost. As a result, only a small number of rig variations exist across most fleets. In this paper, we study how AV perception models are affected by changes in camera viewpoint and propose a way to scale them across vehicle types without repeated data collection and labeling. Using bird's eye view (BEV) segmentation as a motivating task, we find through extensive experiments that existing perception models are surprisingly sensitive to changes in camera viewpoint. When trained with data from one camera rig, small changes to pitch, yaw, depth, or height of the camera at inference time lead to large drops in performance. We introduce a technique for novel view synthesis and use it to transform collected data to the viewpoint of target rigs, allowing us to train BEV segmentation models for diverse target rigs without any additional data collection or labeling cost. To analyze the impact of viewpoint changes, we leverage synthetic data to mitigate other gaps (content, ISP, etc). Our approach is then trained on real data and evaluated on synthetic data, enabling evaluation on diverse target rigs. We release all data for use in future work. Our method is able to recover an average of 14.7% of the IoU that is otherwise lost when deploying to new rigs.

Generalizable face anti-spoofing (FAS) based on domain generalization (DG) has gained growing attention due to its robustness in real-world applications. However, these DG methods rely heavily on labeled source data, which are usually costly and hard to access. Comparably, unlabeled face data are far more accessible in various scenarios. In this paper, we propose the first Unsupervised Domain Generalization framework for Face Anti-Spoofing, namely UDG-FAS, which could exploit large amounts of easily accessible unlabeled data to learn generalizable features for enhancing the low-data regime of FAS. Yet without supervision signals, learning intrinsic live/spoof features from complicated facial information is challenging, which is even tougher in cross-domain scenarios due to domain shift. Existing unsupervised learning methods tend to learn identity-biased and domain-biased features as shortcuts, and fail to specify spoof cues. To this end, we propose a novel Split-Rotation-Merge module to build identity-agnostic local representations for mining intrinsic spoof cues and search the nearest neighbors in the same domain as positives for mitigating the identity bias. Moreover, we propose to search cross-domain neighbors with domain-specific normalization and merged local features to learn a domain-invariant feature space. To our best knowledge, this is the first attempt to learn generalized FAS features in a fully unsupervised way. Extensive experiments show that UDG-FAS significantly outperforms state-of-the-art methods on six diverse practical protocols.

Contemporary LiDAR-based 3D object detection methods mostly focus on single-domain learning or cross-domain adaptive learning. However, for autonomous driving systems, optimizing a specific LiDAR-based 3D object detector for each domain is costly and lacks of scalability in real-world deployment. It is desirable to train a universal LiDAR-based 3D object detector from multiple domains. In this work, we propose the first attempt to explore multi-domain learning and generalization for LiDAR-based 3D object detection. We show that jointly optimizing a 3D object detector from multiple domains achieves better generalization capability compared to the conventional single-domain learning model. To explore informative knowledge across domains towards a universal 3D object detector, we propose a multi-domain knowledge transfer framework with universal feature transformation. This approach leverages spatial-wise and channel-wise knowledge across domains to learn universal feature representations, so it facilitates to optimize a universal 3D object detector for deployment at different domains. Extensive experiments on four benchmark datasets (Waymo, KITTI, NuScenes and ONCE) show the superiority of our approach over the state-of-the-art approaches for multi-domain learning and generalization in LiDAR-based 3D object detection.

Fine-grained and instance-level recognition methods are commonly trained and evaluated on specific domains, in a model per domain scenario. Such an approach, however, is impractical in real large-scale applications. In this work, we address the problem of universal image embedding, where a single universal model is trained and used in multiple domains. First, we leverage existing domain-specific datasets to carefully construct a new large-scale public benchmark for the evaluation of universal image embeddings, with 241k query images, 1.4M index images and 2.8M training images across 8 different domains and 349k classes. We define suitable metrics, training and evaluation protocols to foster future research in this area. Second, we provide a comprehensive experimental evaluation on the new dataset, demonstrating that existing approaches and simplistic extensions lead to worse performance than an assembly of models trained for each domain separately. Finally, we conducted a public research competition on this topic, leveraging industrial datasets, which attracted the participation of more than 1k teams worldwide. This exercise generated many interesting research ideas and findings which we present in detail. Project webpage: https://cmp.felk.cvut.cz/univ_emb/

Medical vision-and-language pre-training (Med-VLP) has shown promising improvements on many downstream medical tasks owing to its applicability to extracting generic representations from medical images and texts. Practically, there exist two typical types, i.e., the fusion-encoder type and the dual-encoder type, depending on whether a heavy fusion module is used. The former is superior at multi-modal tasks owing to the sufficient interaction between modalities; the latter is good at uni-modal and cross-modal tasks due to the single-modality encoding ability. To take advantage of these two types, we propose an effective yet straightforward scheme named PTUnifier to unify the two types. We first unify the input format by introducing visual and textual prompts, which serve as DETR-like queries that assist in extracting features when one of the modalities is missing. By doing so, a single model could serve as a foundation model that processes various tasks adopting different input formats (i.e., image-only, text-only, and image-text-pair). Furthermore, we construct a prompt pool (instead of static ones) to improve diversity and scalability, enabling queries conditioned on different input instances. Experimental results show that our approach achieves state-of-the-art results on a broad range of tasks, spanning uni-modal tasks (i.e., image/text classification and text summarization), cross-modal tasks (i.e., image-to-text generation and image-text/text-image retrieval), and multi-modal tasks (i.e., visual question answering), demonstrating the effectiveness of our approach. Note that the adoption of prompts is orthogonal to most existing Med-VLP approaches and could be a beneficial and complementary extension to these approaches. The source code is available at https://anonymous.4open.science/r/ICCV-2023-Submission-PTUnifier/ and will be released in the final version of this paper.