ICCV 2023 Papers

Understanding and analyzing human behaviors (actions and interactions of people), voices, and sounds in chaotic events is crucial in many applications, e.g., crowd management, emergency response services. Different from human behaviors in daily life, human behaviors in chaotic events are generally different in how they behave and influence others, and hence are often much more complex. However, currently there is lack of a large video dataset for analyzing human behaviors in chaotic situations. To this end, we create the first large and challenging multi-modal dataset, Chaotic World, that simultaneously provides different levels of fine-grained and dense spatio-temporal annotations of sounds, individual actions and group interaction graphs, and even text descriptions for each scene in each video, thereby enabling a thorough analysis of complicated behaviors in crowds and chaos. Our dataset consists of a total of 299,923 annotated instances for detecting human behaviors for Spatiotemporal Action Localization in chaotic events, 224,275 instances for identifying interactions between people for Behavior Graph Analysis in chaotic events, 336,390 instances for localizing relevant scenes of interest in long videos for Spatiotemporal Event Grounding, and 378,093 instances for triangulating the source of sound for Event Sound Source Localization. Given the practical complexity and challenges in chaotic events (e.g., large crowds, serious occlusions, complicated interaction patterns), our dataset shall be able to facilitate the community to develop, adapt, and evaluate various types of advanced models for analyzing human behaviors in chaotic events. We also design a simple yet effective IntelliCare model with a Dynamic Knowledge Pathfinder module that intelligently learns from multiple tasks and can analyze various aspects of a chaotic scene in a unified architecture. This method achieves promising results in experiments. Dataset and code can be found at https://github.com/sutdcv/Chaotic-World

In this paper, we propose a novel center-based decoupled point cloud registration framework for robust 6D object pose estimation in real-world scenarios. Our method decouples the translation from the entire transformation by predicting the object center and estimating the rotation in a center-aware manner. This center offset-based translation estimation is correspondence-free, freeing us from the difficulty of constructing correspondences in challenging scenarios, thus improving robustness. To obtain reliable center predictions, we use a multi-view (bird's eye view and front view) object shape description of the source-point features, with both views jointly voting for the object center. Additionally, we propose an effective shape embedding module to augment the source features, largely completing the missing shape information due to partial scanning, thus facilitating the center prediction. With the center-aligned source and model point clouds, the rotation predictor utilizes feature similarity to establish putative correspondences for SVD-based rotation estimation. In particular, we introduce a center-aware hybrid feature descriptor with a normal correction technique to extract discriminative, part-aware features for high-quality correspondence construction. Our experiments show that our method outperforms the state-of-the-art methods by a large margin on real-world datasets such as TUD-L, LINEMOD, and Occluded-LINEMOD.

Model quantization, which aims to compress deep neural networks and accelerate inference speed, has greatly facilitated the development of cumbersome models on mobile and edge devices. There is a common assumption in quantization methods from prior works that training data is available. In practice, however, this assumption cannot always be fulfilled due to reasons of privacy and security, rendering these methods inapplicable in real-life situations. Thus, data-free network quantization has recently received significant attention in neural network compression. Causal reasoning provides an intuitive way to model casual relationships to eliminate data-driven correlations, making causality an essential component of analyzing data-free problems. However, causal formulations of data-free quantization are inadequate in the literature. To bridge this gap, we construct a causal graph to model the data generation and discrepancy reduction between the pre-trained and quantized models. Inspired by the causal understanding, we propose the Causality-guided Data-free Network Quantization method, Causal-DFQ, to eliminate the reliance on data via approaching an equilibrium of causality-driven intervened distributions. Specifically, we design a content-style-decoupled generator, synthesizing images conditioned on the relevant and irrelevant factors; then we propose a discrepancy reduction loss to align the intervened distributions of the pre-trained and quantized models. It is worth noting that our work is the first attempt towards introducing causality to data-free quantization problem. Extensive experiments demonstrate the efficacy of Causal-DFQ.

Semi-supervised learning (SSL) has recently demonstrated great success in medical image segmentation, significantly enhancing data efficiency with limited annotations. However, despite its empirical benefits, there are still concerns in the literature about the theoretical foundation and explanation of semi-supervised segmentation. To explore this problem, this study first proposes a novel causal diagram to provide a theoretical foundation for the mainstream semi-supervised segmentation methods. Our causal diagram takes two additional intermediate variables into account, which are neglected in previous work. Drawing from this proposed causal diagram, we then introduce a causality-inspired SSL approach on top of co-training frameworks called CauSSL, to improve SSL for medical image segmentation. Specifically, we first point out the importance of algorithmic independence between two networks or branches in SSL, which is often overlooked in the literature. We then propose a novel statistical quantification of the uncomputable algorithmic independence and further enhance the independence via a min-max optimization process. Our method can be flexibly incorporated into different existing SSL methods to improve their performance. Our method has been evaluated on three challenging medical image segmentation tasks using both 2D and 3D network architectures and has shown consistent improvements over state-of-the-art methods. Our code is publicly available at: https://github.com/JuzhengMiao/CauSSL.

Weakly supervised object localization (WSOL) aims to localize objects based on only image-level labels as supervision. Recently, transformers have been introduced into WSOL, yielding impressive results. The self-attention mechanism and multilayer perceptron structure in transformers preserve long-range feature dependency, facilitating complete localization of the full object extent. However, current transformer-based methods predict bounding boxes using category-agnostic attention maps, which may lead to confused and noisy object localization. To address this issue, we propose a novel Category-aware Allocation TRansformer (CATR) that learns category-aware representations for specific objects and produces corresponding category-aware attention maps for object localization. First, we introduce a Category-aware Stimulation Module (CSM) to induce learnable category biases for self-attention maps, providing auxiliary supervision to guide the learning of more effective transformer representations. Second, we design an Object Constraint Module (OCM) to refine the object regions for the category-aware attention maps in a self-supervised manner. Extensive experiments on the CUB-200-2011 and ILSVRC datasets demonstrate that the proposed CATR achieves significant and consistent performance improvements over competing approaches.

Object localization in general environments is a fundamental part of vision systems. While dominating on the COCO benchmark, recent Transformer-based detection methods are not competitive in diverse domains. Moreover, these methods still struggle to very accurately estimate the object bounding boxes in complex environments. We introduce Cascade-DETR for high-quality universal object detection. We jointly tackle the generalization to diverse domains and localization accuracy by proposing the Cascade Attention layer, which explicitly integrates object-centric information into the detection decoder by limiting the attention to the previous box prediction. To further enhance accuracy, we also revisit the scoring of queries. Instead of relying on classification scores, we predict the expected IoU of the query, leading to substantially more well-calibrated confidences. Lastly, we introduce a universal object detection benchmark, UDB10, that contains 10 datasets from diverse domains. While also advancing the state-of-the-art on COCO, Cascade-DETR substantially improves DETR-based detectors on all datasets in UDB10, even by over 10 mAP in some cases. The improvements under stringent quality requirements are even more pronounced. Our code and pretrained models are at https://github.com/SysCV/cascade-detr.

Factored feature volumes offer a simple way to build more compact, efficient, and intepretable neural fields, but also introduce biases that are not necessarily beneficial for real-world data. In this work, we (1) characterize the undesirable biases that these architectures have for axis-aligned signals -- they can lead to radiance field reconstruction differences of as high as 2 PSNR -- and (2) explore how learning a set of canonicalizing transformations can improve representations by removing these biases. We prove in a simple two-dimensional model problem that a hybrid architecture that simultaneously learns these transformations together with scene appearance succeeds with drastically improved efficiency. We validate the resulting architectures, which we call TILTED, using 2D image, signed distance field, and radiance field reconstruction tasks, where we observe improvements across quality, robustness, compactness, and runtime. Results demonstrate that TILTED can enable capabilities comparable to baselines that are 2x larger, while highlighting weaknesses of standard procedures for evaluating neural field representations.

In partial-label learning (PLL), each training example has a set of candidate labels, among which only one is the true label. Most existing PLL studies focus on the instance-independent (II) case, where the generation of candidate labels is only dependent on the true label. However, this II-PLL paradigm could be unrealistic, since candidate labels are usually generated according to the specific features of the instance. Therefore, instance-dependent PLL (ID-PLL) has attracted increasing attention recently. Unfortunately, existing ID-PLL studies lack an insightful perception of the intrinsic challenge in ID-PLL. In this paper, we start with an empirical study of the dynamics of label disambiguation in both II-PLL and ID-PLL. We found that the performance degradation of ID-PLL stems from the inaccurate supervision caused by massive under-disambiguated (UD) examples that do not achieve complete disambiguation. To solve this problem, we propose a novel two-stage PLL framework including selective disambiguation and candidate-aware thresholding. Specifically, we first choose a part of well-disambiguated (WD) examples based on the magnitude of normalized entropy (NE) and integrate harmless complementary supervision from the remaining ones to train two networks. Next, the remaining examples whose NE is lower than the specific class-wise WD-NE threshold are selected as additional WD ones. Meanwhile, the remaining UD examples, whose NE is lower than the self-adaptive UD-NE threshold and whose predictions from two networks are agreed, are also regarded as WD ones for model training. Extensive experiments demonstrate that our proposed method outperforms state-of-the-art PLL methods.

Human readers or radiologists routinely perform full-body multi-organ multi-disease detection and diagnosis in clinical practice, while most medical AI systems are built to focus on single organs with a narrow list of a few diseases. This might severely limit AI's clinical adoption. A certain number of AI models need to be assembled non-trivially to match the diagnostic process of a human reading a CT scan. In this paper, we construct a Unified Tumor Transformer (CancerUniT) model to jointly detect tumor existence & location and diagnose tumor characteristics for eight major cancers in CT scans. CancerUniT is a query-based Mask Transformer model with the output of multi-tumor prediction. We decouple the object queries into organ queries, tumor detection queries and tumor diagnosis queries, and further establish hierarchical relationships among the three groups. This clinically-inspired architecture effectively assists inter- and intra-organ representation learning of tumors and facilitates the resolution of these complex, anatomically related multi-organ cancer image reading tasks. CancerUniT is trained end-to-end using curated large-scale CT images of 10,042 patients including eight major types of cancers and occurring non-cancer tumors (all are pathology-confirmed with 3D tumor masks annotated by radiologists). On the test set of 631 patients, CancerUniT has demonstrated strong performance under a set of clinically relevant evaluation metrics, substantially outperforming both multi-disease methods and an assembly of eight single-organ expert models in tumor detection, segmentation, and diagnosis. This moves one step closer towards a universal high performance cancer screening tool.

We present a framework for generating appropriate facial responses from a listener in dyadic social interactions based on the speaker's words. Given an input transcription of the speaker's words with their timestamps, our approach autoregressively predicts a response of a listener: a sequence of listener facial gestures, quantized using a VQ-VAE. Since gesture is a language component, we propose treating the quantized atomic motion elements as additional language token inputs to a transformer-based large language model. Initializing our transformer with the weights of a language model pre-trained only on text results in significantly higher quality listener responses than training a transformer from scratch. We show that our generated listener motion is fluent and reflective of language semantics through quantitative metrics and a qualitative user study. In our evaluation, we analyze the model's ability to utilize temporal and semantic aspects of spoken text.

We present a novel unsupervised domain adaption method for person re-identification (reID) that generalizes a model trained on a labeled source domain to an unlabeled target domain. We introduce a camera-driven curriculum learning (CaCL) framework that leverages camera labels of person images to transfer knowledge from source to target domains progressively. To this end, we divide target domain dataset into multiple subsets based on the camera labels, and initially train our model with a single subset (i.e., images captured by a single camera). We then gradually exploit more subsets for training, according to a curriculum sequence obtained with a camera-driven scheduling rule. The scheduler considers maximum mean discrepancies (MMD) between each subset and the source domain dataset, such that the subset closer to the source domain is exploited earlier within the curriculum. For each curriculum sequence, we generate pseudo labels of person images in a target domain to train a reID model in a supervised way. We have observed that the pseudo labels are highly biased toward cameras, suggesting that person images obtained from the same camera are likely to have the same pseudo labels, even for different IDs. To address the camera bias problem, we also introduce a camera-diversity (CD) loss encouraging person images of the same pseudo label, but captured across various cameras, to involve more for discriminative feature learning, providing person representations robust to inter-camera variations. Experimental results on standard benchmarks, including real-to-real and synthetic-to-real scenarios, demonstrate the effectiveness of our framework.

Semi-supervised crowd counting is an important yet challenging task. A popular approach is to iteratively generate pseudo-labels for unlabeled data and add them to the training set. The key is to use uncertainty to select reliable pseudo-labels. In this paper, we propose a novel method to calibrate model uncertainty for crowd counting. Our method takes a supervised uncertainty estimation strategy to train the model through a surrogate function. This ensures the uncertainty is well controlled throughout the training. We propose a matching-based patch-wise surrogate function to better approximate uncertainty for crowd counting tasks. The proposed method pays a sufficient amount of attention to details, while maintaining a proper granularity. Altogether our method is able to generate reliable uncertainty estimation, high quality pseudolabels, and achieve state-of-the-art performance in semisupervised crowd counting.

The absolute depth values of surrounding environments provide crucial cues for various assistive technologies, such as localization, navigation, and 3D structure estimation. We propose that accurate depth estimated from panoramic images can serve as a powerful and light-weight input for a wide range of downstream tasks requiring 3D information. While panoramic images can easily capture the surrounding context from commodity devices, the estimated depth shares the limitations of conventional image-based depth estimation; the performance deteriorates under large domain shifts and the absolute values are still ambiguous to infer from 2D observations. By taking advantage of the holistic view, we mitigate such effects in a self-supervised way and fine-tune the network with geometric consistency during the test phase. Specifically, we construct a 3D point cloud from the current depth prediction and project the point cloud at various viewpoints or apply stretches on the current input image to generate synthetic panoramas. Then we minimize the discrepancy of the 3D structure estimated from synthetic images without collecting additional data. We empirically evaluate our method in robot navigation and map-free localization where our method shows large performance enhancements. Our calibration method can therefore widen the applicability under various external conditions, serving as a key component for practical panorama-based machine vision systems.

We present CaPhy, a novel method for reconstructing animatable human avatars with realistic dynamic properties for clothing. Specifically, we aim for capturing the geometric and physical properties of the clothing from real observations. This allows us to apply novel poses to the human avatar with physically correct deformations and wrinkles of the clothing. To this end, we combine unsupervised training with physics-based losses and 3D-supervised training using scanned data to reconstruct a dynamic model of clothing that is physically realistic and conforms to the human scans. We also optimize the physical parameters of the underlying physical model from the scans by introducing gradient constraints of the physics-based losses. In contrast to previous work on 3D avatar reconstruction, our method is able to generalize to novel poses with realistic dynamic cloth deformations. Experiments on several subjects demonstrate that our method can estimate the physical properties of the garments, resulting in superior quantitative and qualitative results compared with previous methods.

Semantic scene completion (SSC) requires an accurate understanding of the geometric and semantic relationships between the objects in the 3D scene for reasoning the occluded objects. The popular SSC methods voxelize the 3D objects, allowing the deep 3D convolutional network (3D CNN) to learn the object relationships from the complex scenes. However, the current networks lack the controllable kernels to model the object relationship across multiple views, where appropriate views provide the relevant information for suggesting the existence of the occluded objects. In this paper, we propose Cross-View Synthesis Transformer (CVSformer), which consists of Multi-View Feature Synthesis and Cross-View Transformer for learning cross-view object relationships. In the multi-view feature synthesis, we use a set of 3D convolutional kernels rotated differently to compute the multi-view features for each voxel. In the cross-view transformer, we employ the cross-view fusion to comprehensively learn the cross-view relationships, which form useful information for enhancing the features of individual views. We use the enhanced features to predict the geometric occupancies and semantic labels of all voxels. We evaluate CVSformer on public datasets, where CVSformer yields state-of-the-art results. Our code is available at https://github.com/donghaotian123/CVSformer.

Recent advances in neural reconstruction using posed image sequences have made remarkable progress. However, due to the lack of depth information, existing volumetric-based techniques simply duplicate 2D image features of the object surface along the entire camera ray. We contend this duplication introduces noise in empty and occluded spaces, posing challenges for producing high-quality 3D geometry. Drawing inspiration from traditional multi-view stereo methods, we propose an end-to-end 3D neural reconstruction framework CVRecon, designed to exploit the rich geometric embedding in the cost volumes to facilitate 3D geometric feature learning. Furthermore, we present Ray-contextual Compensated Cost Volume (RCCV), a novel 3D geometric feature representation that encodes view-dependent information with improved integrity and robustness. Through comprehensive experiments, we demonstrate that our approach significantly improves the reconstruction quality in various metrics and recovers clear fine details of the 3D geometries. Our extensive ablation studies provide insights into the development of effective 3D geometric feature learning schemes. Project page: https://cvrecon.ziyue.cool

The discrimination of instance embeddings plays a vital role in associating instances across time for online video instance segmentation (VIS). Instance embedding learning is directly supervised by the contrastive loss computed upon the contrastive items (CIs), which are sets of anchor/positive/negative embeddings. Recent online VIS methods leverage CIs sourced from one reference frame only, which we argue is insufficient for learning highly discriminative embeddings. Intuitively, a possible strategy to enhance CIs is replicating the inference phase during training. To this end, we propose a simple yet effective training strategy, called Consistent Training for Online VIS(CTVIS), which devotes to aligning the training and inference pipelines in terms of building CIs. Specifically, CTVIS constructs CIs by referring inference the momentum-averaged embedding and the memory bank storage mechanisms, and adding noise to the relevant embeddings. Such an extension allows a reliable comparison between embeddings of current instances and the stable representations of historical instances, thereby conferring an advantage in modeling VIS challenges such as occlusion, re-identification, and deformation. Empirically, CTVIS outstrips the SOTA VIS models by up to +5.0 points on three VIS benchmarks, including YTVIS19 (55.1% AP), YTVIS21 (50.1% AP) and OVIS (35.5% AP). Furthermore, we find that pseudo-videos transformed from images can train robust models surpassing fully-supervised ones.

Vision-Language Pretraining (VLP) has shown impressive results on diverse downstream tasks by offline training on large-scale datasets. Regarding the growing nature of real-world data, such an offline training paradigm on ever-expanding data is unsustainable, because models lack the continual learning ability to accumulate knowledge constantly. However, most continual learning studies are limited to uni-modal classification and existing multi-modal datasets cannot simulate continual non-stationary data stream scenarios. To support the study of Vision-Language Continual Pretraining (VLCP), we first contribute a comprehensive and unified benchmark dataset P9D which contains over one million product image-text pairs from 9 industries. The data from each industry as an independent task supports continual learning and conforms to the real-world long-tail nature to simulate pretraining on web data. We comprehensively study the characteristics and challenges of VLCP, and propose a new algorithm: Compatible momentum contrast with Topology Preservation, dubbed CTP. The compatible momentum model absorbs the knowledge of the current and previous-task models to flexibly update the modal feature. Moreover, Topology Preservation transfers the knowledge of embedding across tasks while preserving the flexibility of feature adjustment. The experimental results demonstrate our method not only achieves superior performance compared with other baselines but also does not bring an expensive training burden. Dataset and codes are available at https://github.com/KevinLight831/CTP.

Domain Adaptive Object Detection (DAOD) aims to improve the detection performance of target domains by minimizing the feature distribution between the source and target domain. Recent approaches usually align such distributions in terms of categories through adversarial learning and some progress has been made. However, when objects are non-uniformly distributed at different scales, such category-level alignment causes imbalanced object feature learning, refer as the inconsistency of category alignment at different scales. For better category-level feature alignment, we propose a novel DAOD framework of joint category and scale information, dubbed CSDA, such a design enables effective object learning for different scales. Specifically, our framework is implemented by two closely-related modules: 1) SGFF (Scale-Guided Feature Fusion) fuses the category representations of different domains to learn category-specific features, where the features are aligned by discriminators at three scales. 2) SAFE (Scale-Auxiliary Feature Enhancement) encodes scale coordinates into a group of tokens and enhances the representation of category-specific features at different scales by self-attention. Based on the anchor-based Faster-RCNN and anchor-free FCOS detectors, experiments show that our method achieves state-of-the-art results on three DAOD benchmarks.