ICCV 2023 Papers

In this paper, we introduce audio-visual class-incremental learning, a class-incremental learning scenario for audio-visual video recognition. We demonstrate that joint audio-visual modeling can improve class-incremental learning, but current methods fail to preserve semantic similarity between audio and visual features as incremental step grows. Furthermore, we observe that audio-visual correlations learned in previous tasks can be forgotten as incremental steps progress, leading to poor performance. To overcome these challenges, we propose AV-CIL, which incorporates Dual-Audio-Visual Similarity Constraint (D-AVSC) to maintain both instance-aware and class-aware semantic similarity between audio-visual modalities and Visual Attention Distillation (VAD) to retain previously learned audio-guided visual attentive ability. We create three audio-visual class-incremental datasets, AVE-Class-Incremental (AVE-CI), Kinetics-Sounds-Class-Incremental (K-S-CI), and VGGSound100-Class-Incremental (VS100-CI) based on the AVE, Kinetics-Sounds, and VGGSound datasets, respectively. Our experiments on AVE-CI, K-S-CI, and VS100-CI demonstrate that AV-CIL significantly outperforms existing class-incremental learning methods in audio-visual class-incremental learning. Code and data are available at: https://github.com/weiguoPian/AV-CIL_ICCV2023.

Text-to-video retrieval systems have recently made significant progress by utilizing pre-trained models trained on large-scale image-text pairs. However, most of the latest methods primarily focus on the video modality while disregarding the audio signal for this task. Nevertheless, a recent advancement by EclipSE has improved long-range text-to-video retrieval by developing an audiovisual video representation. Nonetheless, the objective of the text-to-video retrieval task is to capture the complementary audio and video information that is pertinent to the text query rather than simply achieving better audio and video alignment. To address this issue, we introduce TEFAL, a TExt-conditioned Feature ALignment method that produces both audio and video representations conditioned on the text query. Instead of using only an audiovisual attention block, which could suppress the audio information relevant to the text query, our approach employs two independent cross-modal attention blocks that enable the text to attend to the audio and video representations separately. Our proposed method's efficacy is demonstrated on four benchmark datasets that include audio: MSR-VTT, LSMDC, VATEX, and Charades, and achieves better than state-of-the-art performance consistently across the four datasets. This is attributed to the additional text-query-conditioned audio representation and the complementary information it adds to the text-query-conditioned video representation.

In vision-language modeling, image token removal is an efficient augmentation technique to reduce the cost of encoding image features. The CLIP-style models, however, have been found to be negatively impacted by this technique. We hypothesize that removing a large portion of image tokens may inadvertently destroy the semantic information associated to a given text description, resulting in misaligned paired data in CLIP training. To address this issue, we propose an attentive token removal approach, which retains a small number of tokens that have a strong semantic correlation to the corresponding text description. The correlation scores are dynamically evaluated through an EMA-updated vision encoder. Our method, termed attentive mask CLIP, outperforms original CLIP and CLIP variant with random token removal while saving the training time. In addition, our approach also enables efficient multi-view contrastive learning. Experimentally, by training ViT-B on YFCC-15M dataset, our approach achieves 43.9% top-1 accuracy on ImageNet-1K zero-shot classification, 62.7/42.1 and 38.0/23.2 I2T/T2I retrieval accuracy on Flickr30K and MS COCO, outperforming SLIP by +1.1%,+5.5/+0.9, and +4.4/+1.3, respectively, while being 2.30x faster. An efficient version of our approach runs 1.16x faster than the plain CLIP model, while achieving significant gains of +5.3%, +11.3/+8.0, and +9.5/+4.9 on these benchmarks, respectively.

We propose a novel end-to-end document understanding model called SeRum (SElective Region Understanding Model) for extracting meaningful information from document images, including document analysis, retrieval, and office automation. Unlike state-of-the-art approaches that rely on multi-stage technical schemes and are computationally expensive, SeRum converts document image understanding and recognition tasks into a local decoding process of the vision tokens of interest, using a content-aware token merge module. This mechanism enables the model to pay more attention to regions of interest generated by the query decoder, improving the model's effectiveness and speeding up the decoding speed of the generative scheme. We also designed several pre-training tasks to enhance the understanding and local awareness of the model. Experimental results demonstrate that SeRum achieves state-of-the-art performance on document understanding tasks and competitive results on text spotting tasks. SeRum represents a substantial advancement towards enabling efficient and effective end-to-end document understanding.

This paper describes an attention-driven approach to 3-D point cloud sampling. We establish our method based on a structure-aware attention discriminant analysis that explores geometric and semantic relations embodied among points and their clusters. The proposed attention discriminant sampling (ADS) starts by efficiently decomposing a given point cloud into clusters to implicitly encode its structural and geometric relatedness among points. By treating each cluster as a structural component, ADS then draws on evaluating two levels of self-attention: within-cluster and between-cluster. The former reflects the semantic complexity entailed by the learned features of points within each cluster, while the latter reveals the semantic similarity between clusters. Driven by structurally preserving the point distribution, these two aspects of self-attention help avoid sampling redundancy and decide the number of sampled points in each cluster. Extensive experiments demonstrate that ADS significantly improves classification performance to 95.1% on ModelNet40 and 87.5% on ScanObjectNN and achieves 86.9% mIoU on ShapeNet Part Segmentation. For scene segmentation, ADS yields 91.1% accuracy on S3DIS with higher mIoU to the state-of-the-art and 75.6% mIoU on ScanNetV2. Furthermore, ADS surpasses the state-of-the-art with 55.0% mAP50 on ScanNetV2 object detection.

Generating 3D human motion based on textual descriptions has been a research focus in recent years. It requires the generated motion to be diverse, natural, and conform to the textual description. Due to the complex spatio-temporal nature of human motion and the difficulty in learning the cross-modal relationship between text and motion, text-driven motion generation is still a challenging problem. To address these issues, we propose AttT2M, a two-stage method with multi-perspective attention mechanism: body-part attention and global-local motion-text attention. The former focuses on the motion embedding perspective, which means introducing a body-part spatio-temporal encoder into VQ-VAE to learn a more expressive discrete latent space. The latter is from the cross-modal perspective, which is used to learn the sentence-level and word-level motion-text cross-modal relationship. The text-driven motion is finally generated with a generative transformer. Extensive experiments conducted on HumanML3D and KIT-ML demonstrate that our method outperforms the current state-of-the-art works in terms of qualitative and quantitative evaluation, and achieve fine-grained synthesis and action2motion. Our code will be publicly available.

Computer vision-based walking assistants are prominent tools for aiding visually impaired people in navigation. Blind road segmentation is a key element in these walking assistant systems. However, most walking assistant systems rely on visual light images, which is dangerous in weak illumination environments such as darkness or fog. To address this issue and enhance the safety of vision-based walking assistant systems, we developed a thermal infrared blind road segmentation neural network (TINN). In contrast to conventional segmentation techniques that primarily concentrate on enhancing feature extraction and perception, our approach is geared towards preserving the inherent radiation characteristics within the thermal imaging process. Initially, we modelled two critical factors in thermal infrared imaging - thermal light atmospheric transmission and thermal inertia effect. Subsequently, we use an encoder-decoder architecture to fuse the feathers extracted by the two modules. Additionally, to train the network and evaluate the effectiveness of the proposed method, we constructed a large-scale thermal infrared blind road segmentation dataset named TBRSD consists 5180 pixel-level manual annotations. The experimental results demonstrate that our method outperforms existing techniques and achieves state-of-the-art performance in thermal blind road segmentation, as validated on benchmark thermal infrared semantic segmentation datasets such as MFNet and SODA. The dataset and our code are both publicly available in https://github.com/chenjzBUAA/TBRSD or http://xzbai.buaa.edu.cn/datasets.html.

Both indoor and outdoor environments are inherently structured and repetitive. Traditional modeling pipelines keep an asset library storing unique object templates, which is both versatile and memory efficient in practice. Inspired by this observation, we propose AssetField, a novel neural scene representation that learns a set of object-aware ground feature planes to represent the scene, where an asset library storing template feature patches can be constructed in an unsupervised manner. Unlike existing methods which require object masks to query spatial points for object editing, our ground feature plane representation offers a natural visualization of the scene in the bird-eye view, allowing a variety of operations (e.g. translation, duplication, deformation) on objects to configure a new scene. With the template feature patches, group editing is enabled for scenes with many recurring items to avoid repetitive work on object individuals. We show that AssetField not only achieves competitive performance for novel-view synthesis but also generates realistic renderings for new scene configurations.

We introduce the Aria Digital Twin (ADT) - an egocentric dataset captured using Aria glasses with extensive object, environment, and human level ground truth. This ADT release contains 200 sequences of real-world activities conducted by Aria wearers in two real indoor scenes with 398 object instances (344 stationary and 74 dynamic). Each sequence consists of: a) raw data of two monochrome camera streams, one RGB camera stream, two IMU streams; b) complete sensor calibration; c) ground truth data including continuous 6-degree-of-freedom (6DoF) poses of the Aria devices, object 6DoF poses, 3D eye gaze vectors, 3D human poses, 2D image segmentations, image depth maps; and d) photo-realistic synthetic renderings. To the best of our knowledge, there is no existing egocentric dataset with a level of accuracy, photo-realism and comprehensiveness comparable to ADT. By contributing ADT to the research community, our mission is to set a new standard for evaluation in the egocentric machine perception domain, which includes very challenging research problems such as 3D object detection and tracking, scene reconstruction and understanding, sim-to-real learning, human pose prediction - while also inspiring new machine perception tasks for augmented reality (AR) applications. To kick start exploration of the ADT research use cases, we evaluated several existing state-of-the-art methods for object detection, segmentation and image translation tasks that demonstrate the usefulness of ADT as a benchmarking dataset.

We demonstrate a compact, cost-effective snapshot spectral imaging system named Aperture Diffraction Imaging Spectrometer (ADIS), which consists only of an imaging lens with an ultra-thin orthogonal aperture mask and a mosaic filter sensor, requiring no additional physical footprint compared to common RGB cameras. Then we introduce a new optical design that each point in the object space is multiplexed to discrete encoding locations on the mosaic filter sensor by diffraction-based spatial-spectral projection engineering generated from the orthogonal mask. The orthogonal projection is uniformly accepted to obtain a weakly calibration-dependent data form to enhance modulation robustness. Meanwhile, the Cascade Shift-Shuffle Spectral Transformer (CSST) with strong perception of the diffraction degeneration is designed to solve a sparsity-constrained inverse problem, realizing the volume reconstruction from 2D measurements with Large amount of aliasing. Our system is evaluated by elaborating the imaging optical theory and reconstruction algorithm with demonstrating the experimental imaging under a single exposure. Ultimately, we achieve the sub-super-pixel spatial resolution and high spectral resolution imaging. The code will be available at: https://github.com/Krito-ex/CSST.

Text-to-image diffusion models are nothing but a revolution, allowing anyone, even without design skills, to create realistic images from simple text inputs. With powerful personalization tools like DreamBooth, they can generate images of a specific person just by learning from his/her few reference images. However, when misused, such a powerful and convenient tool can produce fake news or disturbing content targeting any individual victim, posing a severe negative social impact. In this paper, we explore a defense system called Anti-DreamBooth against such malicious use of DreamBooth. The system aims to add subtle noise perturbation to each user's image before publishing in order to disrupt the generation quality of any DreamBooth model trained on these perturbed images. We investigate a wide range of algorithms for perturbation optimization and extensively evaluate them on two facial datasets over various text-to-image model versions. Despite the complicated formulation of DreamBooth and Diffusion-based text-to-image models, our methods effectively defend users from the malicious use of those models. Their effectiveness withstands even adverse conditions, such as model or prompt/term mismatching between training and testing. Our code will be available at https://github.com/VinAIResearch/Anti-DreamBooth

Unsupervised anomaly detection is widely studied for industrial applications since it is difficult to obtain anomalous data. In particular, reconstruction-based anomaly detection can be a feasible solution if there is no option to use external knowledge, such as extra datasets or pre-trained models. However, reconstruction-based methods have limited utility due to poor detection performance. A score-based model, also known as a denoising diffusion model, recently has shown a high sample quality in the generation task. In this paper, we propose a novel unsupervised anomaly detection method leveraging the score-based model. This method promises good performance without external knowledge. The score, a gradient of the log-likelihood, has a property that is available for anomaly detection. The samples on the data manifold can be restored instantly by the score, even if they are randomly perturbed. We call this a score-based perturbation resilience. On the other hand, the samples that deviate from the manifold cannot be restored in the same way. The variation of resilience depending on the sample position can be an indicator to discriminate anomalies. We derive this statement from a geometric perspective. Our method shows superior performance on three benchmark datasets for industrial anomaly detection. Specifically, on MVTec AD, we achieve image-level AUROC of 97.7% and pixel-level AUROC of 97.4% outperforming previous works that do not use external knowledge.

Anomaly detection (AD) is a crucial machine learning task that aims to learn patterns from a set of normal training samples to identify abnormal samples in test data. Most existing AD studies assume that the training and test data are drawn from the same data distribution, but the test data can have large distribution shifts arising in many real-world applications due to different natural variations such as new lighting conditions, object poses, or background appearances, rendering existing AD methods ineffective in such cases. In this paper, we consider the problem of anomaly detection under distribution shift and establish performance benchmarks on four widely-used AD and out-of-distribution (OOD) generalization datasets. We demonstrate that simple adaptation of state-of-the-art OOD generalization methods to AD settings fails to work effectively due to the lack of labeled anomaly data. We further introduce a novel robust AD approach to diverse distribution shifts by minimizing the distribution gap between in-distribution and OOD normal samples in both the training and inference stages in an unsupervised way. Our extensive empirical results on the four datasets show that our approach substantially outperforms state-of-the-art AD methods and OOD generalization methods on data with various distribution shifts, while maintaining the detection accuracy on in-distribution data. Code and data are available at https://github.com/mala-lab/ADShift.

Accurately estimating the 3D pose and shape is an essential step towards understanding animal behavior, and can potentially benefit many downstream applications, such as wildlife conservation. However, research in this area is held back by the lack of a comprehensive and diverse dataset with high-quality 3D pose and shape annotations. In this paper, we propose Animal3D, the first comprehensive dataset for mammal animal 3D pose and shape estimation. Animal3D consists of 3379 images collected from 40 mammal species, high-quality annotations of 26 keypoints, and importantly the pose and shape parameters of the SMAL model. All annotations were labeled and checked manually in a multi-stage process to ensure highest quality results. Based on the Animal3D dataset, we benchmark representative shape and pose estimation models at: (1) supervised learning from only the Animal3D data, (2) synthetic to real transfer from synthetically generated images, and (3) fine-tuning human pose and shape estimation models. Our experimental results demonstrate that predicting the 3D shape and pose of animals across species remains a very challenging task, despite significant advances in human pose estimation and animal pose estimation for specific species. Our results further demonstrate that synthetic pre-training is a viable strategy to boost the model performance. Overall, Animal3D opens new directions for facilitating future research in animal 3D pose and shape estimation, and is publicly available.

Anchor-based detectors have been continuously developed for object detection. However, the individual anchor box makes it difficult to predict the boundary's offset accurately. Instead of taking each bounding box as a closed individual, we consider using multiple boxes together to get prediction boxes. To this end, this paper proposes the Box Decouple-Couple(BDC) strategy in the inference, which no longer discards the overlapping boxes, but decouples the corner points of these boxes. Then, according to each corner's score, we couple the corner points to select the most accurate corner pairs. To meet the BDC strategy, a simple but novel model is designed named the Anchor-Intermediate Detector(AID), which contains two head networks, i.e., an anchor-based head and an anchor-free Corner-aware head. The corner-aware head is able to score the corners of each bounding box to facilitate the coupling between corner points. Extensive experiments on MS COCO show that the proposed anchor-intermediate detector respectively outperforms their baseline RetinaNet and GFL method by 2.4 and 1.2 AP on the MS COCO test-dev dataset without any bells and whistles.

The tensor-based multi-view subspace clustering algorithms have received widespread attention due to the powerful ability to capture high-order correlation across views. Although such algorithms have achieved remarkable success, they still suffer from three main issues: 1) The extremely high computational complexity makes tensor-based methods difficult to handle large-scale data sets. 2) The commonly used Tensor Nuclear Norm (TNN) treats different singular values equally and under-penalizes the noise components, resulting in a sub-optimal representation tensor. 3) The subspace-based methods usually ignore the local geometric structure of the original data. Being aware of these, we propose Anchor Structure Regularitation Induced Multi-view Subspace Clustering via Enhanced Tensor Rank Minimization (ASR-ETR). Specifically, an anchor representation tensor is constructed by using the anchor representation strategy rather than the self-representation strategy to reduce the time complexity, and an Anchor Structure Regularization (ASR) is employed to enhance the local geometric structure in the learned anchor-representation tensor. We further define an Enhanced Tensor Rank (ETR), which is a tighter surrogate of the tensor rank and more effective to drive the noise out. Moreover, an efficient iterative optimization algorithm is designed to solve the ASR-ETR, which enjoys both linear complexity and favorable convergence. Extensive experimental results on various data sets demonstrate the superiority of the proposed algorithm as compared to state-of-the-art methods.

Self-supervised learning (SSL) has recently achieved promising performance for 3D medical image analysis tasks. Most current methods follow existing SSL paradigm originally designed for photographic or natural images, which cannot explicitly and thoroughly exploit the intrinsic similar anatomical structures across varying medical images. This may in fact degrade the quality of learned deep representations by maximizing the similarity among features containing spatial misalignment information and different anatomical semantics. In this work, we propose a new self-supervised learning framework, namely Alice, that explicitly fulfills Anatomical invariance modeling and semantic alignment via elaborately combining discriminative and generative objectives. Alice introduces a new contrastive learning strategy which encourages the similarity between views that are diversely mined but with consistent high-level semantics, in order to learn invariant anatomical features. Moreover, we design a conditional anatomical feature alignment module to complement corrupted embeddings with globally matched semantics and inter-patch topology information, conditioned by the distribution of local image content, which permits to create better contrastive pairs. Our extensive quantitative experiments on three 3D medical image analysis tasks demonstrate and validate the performance superiority of Alice, surpassing the previous best SSL counterpart methods and showing promising ability for united representation learning. Codes are available at https://github.com/alibaba-damo-academy/alice.

As a new paradigm in machine learning, self-supervised learning (SSL) is capable of learning high-quality representations of complex data without relying on labels. In addition to eliminating the need for labeled data, research has found that SSL improves the adversarial robustness over supervised learning since lacking labels makes it more challenging for adversaries to manipulate model predictions. However, the extent to which this robustness superiority generalizes to other types of attacks remains an open question. We explore this question in the context of backdoor attacks. Specifically, we design and evaluate CTRL, an embarrassingly simple yet highly effective self-supervised backdoor attack. By only polluting a tiny fraction of training data (<1%) with indistinguishable poisoning samples, CTRL causes any trigger-embedded input to be misclassified to the adversary's designated class with a high probability (>99%) at inference time. Our findings suggest that SSL and supervised learning are comparably vulnerable to backdoor attacks. More importantly, through the lens of CTRL, we study the inherent vulnerability of SSL to backdoor attacks. With both empirical and analytical evidence, we reveal that the representation invariance property of SSL, which benefits adversarial robustness, may also be the very reason making SSL highly susceptible to backdoor attacks. Our findings also imply that the existing defenses against supervised backdoor attacks are not easily retrofitted to the unique vulnerability of SSL.

While the transferability property of adversarial examples allows the adversary to perform black-box attacks i.e., the attacker has no knowledge about the target model), the transfer-based adversarial attacks have gained great attention. Previous works mostly study gradient variation or image transformations to amplify the distortion on critical parts of inputs. These methods can work on transferring across models with limited differences, i.e., from CNNs to CNNs, but always fail in transferring across models with wide differences, such as from CNNs to ViTs. Alternatively, model ensemble adversarial attacks are proposed to fuse outputs from surrogate models with diverse architectures to get an ensemble loss, making the generated adversarial example more likely to transfer to other models as it can fool multiple models concurrently. However, existing ensemble attacks simply fuse the outputs of the surrogate models evenly, thus are not efficacious to capture and amplify the intrinsic transfer information of adversarial examples. In this paper, we propose an adaptive ensemble attack, dubbed AdaEA, to adaptively control the fusion of the outputs from each model, via monitoring the discrepancy ratio of their contributions towards the adversarial objective. Furthermore, an extra disparity-reduced filter is introduced to further synchronize the update direction. As a result, we achieve considerable improvement over the existing ensemble attacks on various datasets, and the proposed AdaEA can also boost existing transfer-based attacks, which further demonstrates its efficacy and versatility.