ICCV 2023 论文列表

We propose a novel transformer-based framework that reconstructs two high fidelity hands from multi-view RGB images. Unlike existing hand pose estimation methods, where one typically trains a deep network to regress hand model parameters from single RGB image, we consider a more challenging problem setting where we directly regress the absolute root poses of two-hands with extended forearm at high resolution from egocentric view. As existing datasets are either infeasible for egocentric viewpoints or lack background variations, we create a large-scale synthetic dataset with diverse scenarios and collect a real dataset from multi-calibrated camera setup to verify our proposed multi-view image feature fusion strategy. To make the reconstruction physically plausible, we propose two strategies: (i) a coarse-to-fine spectral graph convolution decoder to smoothen the meshes during upsampling and (ii) an optimisation-based refinement stage at inference to prevent self-penetrations. Through extensive quantitative and qualitative evaluations, we show that our framework is able to produce realistic two-hand reconstructions and demonstrate the generalisation of synthetic-trained models to real data, as well as real-time AR/VR applications.

The frame quality deterioration problem is one of the main challenges in the field of video understanding. To compensate for the information loss caused by deteriorated frames, recent approaches exploit transformer-based integration modules to obtain spatio-temporal information. However, these integration modules are heavy and complex. Furthermore, each integration module is specifically tailored for its target task, making it difficult to generalise to multiple tasks. In this paper, we present a neat and unified framework, called Spatio-Temporal Prompting Network (STPN). It can efficiently extract robust and accurate video features by dynamically adjusting the input features in the backbone network. Specifically, STPN predicts several video prompts containing spatio-temporal information of neighbour frames. Then, these video prompts are prepended to the patch embeddings of the current frame as the updated input for video feature extraction. Moreover, STPN is easy to generalise to various video tasks because it does not contain task-specific modules. Without bells and whistles, STPN achieves state-of-the-art performance on three widely-used datasets for different video understanding tasks, i.e., ImageNetVID for video object detection, YouTubeVIS for video instance segmentation, and GOT-10k for visual object tracking. Codes are available at https://github.com/guanxiongsun/STPN.

Multi-agent collaborative perception as a potential application for vehicle-to-everything communication could significantly improve the perception performance of autonomous vehicles over single-agent perception. However, several challenges remain in achieving pragmatic information sharing in this emerging research. In this paper, we propose SCOPE, a novel collaborative perception framework that aggregates the spatio-temporal awareness characteristics across on-road agents in an end-to-end manner. Specifically, SCOPE has three distinct strengths: i) it considers effective semantic cues of the temporal context to enhance current representations of the target agent; ii) it aggregates perceptually critical spatial information from heterogeneous agents and overcomes localization errors via multi-scale feature interactions; iii) it integrates multi-source representations of the target agent based on their complementary contributions by an adaptive fusion paradigm. To thoroughly evaluate SCOPE, we consider both real-world and simulated scenarios of collaborative 3D object detection tasks on three datasets. Extensive experiments show the superiority of our approach and the necessity of the proposed components. The project link is https://ydk122024.github.io/SCOPE/.

We propose Spatio-temporal Crop Aggregation for video representation LEarning (SCALE), a novel method that enjoys high scalability at both training and inference time. Our model builds long-range video features by learning from sets of video clip-level features extracted with a pre-trained backbone. To train the model, we propose a self-supervised objective consisting of masked clip feature predictions. We apply sparsity to both the input, by extracting a random set of video clips, and to the loss function, by only reconstructing the sparse inputs. Moreover, we use dimensionality reduction by working in the latent space of a pre-trained backbone applied to single video clips. These techniques make our method not only extremely efficient to train but also highly effective in transfer learning. We demonstrate that our video representation yields state-of-the-art performance with linear, nonlinear, and k-NN probing on common action classification and video understanding datasets.

Although deep learning-based solutions have achieved impressive reconstruction performance in image super-resolution (SR), these models are generally large, with complex architectures, making them incompatible with low-power devices with many computational and memory constraints. To overcome these challenges, we propose a spatially-adaptive feature modulation (SAFM) mechanism for efficient SR design. In detail, the SAFM layer uses independent computations to learn multi-scale feature representations and aggregates these features for dynamic spatial modulation. As the SAFM prioritizes exploiting non-local feature dependencies, we further introduce a convolutional channel mixer (CCM) to encode local contextual information and mix channels simultaneously. Extensive experimental results show that the proposed method is 3x smaller than state-of-the-art efficient SR methods, e.g., IMDN, and yields comparable performance with much less memory usage. Our source codes and pre-trained models are available at: https://github.com/sunny2109/SAFMN.

Cycle consistency has long been exploited as a powerful prior for jointly optimizing maps within a collection of shapes. In this paper, we investigate its utility in the approaches of Deep Functional Maps, which are considered state-of-the-art in non-rigid shape matching. We first justify that under certain conditions, the learned maps, when represented in the spectral domain, are already cycle consistent. Furthermore, we identify the discrepancy that spectrally consistent maps are not necessarily spatially, or point-wise, consistent. In light of this, we present a novel design of unsupervised Deep Functional Maps, which effectively enforces the harmony of learned maps under the spectral and the point-wise representation. By taking advantage of cycle consistency, our framework produces state-of-the-art results in mapping shapes even under significant distortions. Beyond that, by independently estimating maps in both spectral and spatial domains, our method naturally alleviates over-fitting in network training, yielding superior generalization performance and accuracy within an array of challenging tests for both near-isometric and non-isometric datasets.

Weakly supervised object localization (WSOL) is a challenging task aiming to localize objects with only image-level supervision. Recent works apply visual transformer to WSOL and achieve significant success by exploiting the long-range feature dependency in self-attention mechanism. However, existing transformer-based methods synthesize the classification feature maps as the localization map, which leads to optimization conflicts between classification and localization tasks. To address this problem, we propose to learn a task-specific spatial-aware token (SAT) to condition localization in a weakly supervised manner. Specifically, a spatial token is first introduced in the input space to aggregate representations for localization task. Then a spatial aware attention module is constructed, which allows spatial token to generate foreground probabilities of different patches by querying and to extract localization knowledge from the classification task. Besides, for the problem of sparse and unbalanced pixel-level supervision obtained from the image-level label, two spatial constraints, including batch area loss and normalization loss, are designed to compensate and enhance this supervision. Experiments show that the proposed SAT achieves state-of-the-art performance on both CUB-200 and ImageNet, with 98.45% and 73.13% GT-known Loc, respectively. Even under the extreme setting of using only 1 image per class from ImageNet for training, SAT already exceeds the SOTA method by 2.1% GT-known Loc. Code and models are available at https://github.com/wpy1999/SAT.

Object detection via inaccurate bounding box supervision has boosted a broad interest due to the expensive high-quality annotation data or the occasional inevitability of low annotation quality (e.g. tiny objects). The previous works usually utilize multiple instance learning (MIL), which highly depends on category information, to select and refine a low-quality box. Those methods suffer from part domination, object drift and group prediction problems without exploring spatial information. In this paper, we heuristically propose a Spatial Self-Distillation based Object Detector (SSD-Det) to mine spatial information to refine the inaccurate box in a self-distillation fashion. SSD-Det utilizes a Spatial Position Self-Distillation SPSD) module to exploit spatial information and an interactive structure to combine spatial information and category information, thus constructing a high-quality proposal bag. To further improve the selection procedure, a Spatial Identity Self-Distillation (SISD) module is introduced in SSD-Det to obtain spatial confidence to help select the best proposals. Experiments on MS-COCO and VOC datasets with noisy box annotation verify our method's effectiveness and achieve state-of-the-art performance. The code is available at https://github.com/ucas-vg/PointTinyBenchmark/tree/SSD-Det.

Neural Radiance Field (NeRF) significantly degrades when only a limited number of views are available. To complement the lack of 3D information, depth-based models, such as DSNeRF and MonoSDF, explicitly assume the availability of accurate depth maps of multiple views. They linearly scale the accurate depth maps as supervision to guide the predicted depth of few-shot NeRFs. However, accurate depth maps are difficult and expensive to capture due to wide-range depth distances in the wild. This work presents a new Sparse-view NeRF (SparseNeRF) framework that exploits depth priors from real-world inaccurate observations. The inaccurate depth observations are either from pre-trained depth models or coarse depth maps of consumer-level depth sensors. Since coarse depth maps are not strictly scaled to the ground-truth depth maps, we propose a simple yet effective constraint, a local depth ranking method, on NeRFs such that the expected depth ranking of the NeRF is consistent with that of the coarse depth maps in local patches. To preserve the spatial continuity of the estimated depth of NeRF, we further propose a spatial continuity constraint to encourage the consistency of the expected depth continuity of NeRF with coarse depth maps. Surprisingly, with simple depth ranking constraints, SparseNeRF outperforms all state-of-the-art few-shot NeRF methods (including depth-based models) on standard LLFF and DTU datasets. Moreover, we collect a new dataset NVS-RGBD that contains real-world depth maps from Azure Kinect, ZED 2, and iPhone 13 Pro. Extensive experiments on NVS-RGBD dataset also validate the superiority and generalizability of SparseNeRF. Code and dataset are available at https://sparsenerf.github.io/.

Masked Autoencoders (MAE) and its variants have proven to be effective for pretraining large-scale Vision Transformers (ViTs). However, small-scale models do not benefit from the pretraining mechanisms due to limited capacity. Sparse training is a method of transferring representations from large models to small ones by pruning unimportant parameters. However, naively combining MAE finetuning with sparse training make the network task-specific, resulting in the loss of task-agnostic knowledge, which is crucial for model generalization. In this paper, we aim to reduce model complexity from large vision transformers pretrained by MAE with assistant of sparse training. We summarize various sparse training methods to prune large vision transformers during MAE pretraining and finetuning stages, and discuss their shortcomings. To improve learning both task-agnostic and task-specific knowledge, we propose SparseMAE, a novel two-stage sparse training method that includes sparse pretraining and sparse finetuning. In sparse pretraining, we dynamically prune a small-scale sub-network from a ViT-Base. During finetuning, the sparse sub-network adaptively changes its topology connections under the task-agnostic knowledge of the full model. Extensive experimental results demonstrate the effectiveness of our method and its superiority on small-scale vision transformers. Code will be available at https://github.com/aojunzz/SparseMAE.

By identifying four important components of existing LiDAR-camera 3D object detection methods (LiDAR and camera candidates, transformation, and fusion outputs), we observe that all existing methods either find dense candidates or yield dense representations of scenes. However, given that objects occupy only a small part of a scene, finding dense candidates and generating dense representations is noisy and inefficient. We propose SparseFusion, a novel multi-sensor 3D detection method that exclusively uses sparse candidates and sparse representations. Specifically, SparseFusion utilizes the outputs of parallel detectors in the LiDAR and camera modalities as sparse candidates for fusion. We transform the camera candidates into the LiDAR coordinate space by disentangling the object representations. Then, we can fuse the multi-modality candidates in a unified 3D space by a lightweight self-attention module. To mitigate negative transfer between modalities, we propose novel semantic and geometric cross-modality transfer modules that are applied prior to the modality-specific detectors. SparseFusion achieves state-of-the-art performance on the nuScenes benchmark while also running at the fastest speed, even outperforming methods with stronger backbones. We perform extensive experiments to demonstrate the effectiveness and efficiency of our modules and overall method pipeline. Our code will be made publicly available at https://github.com/yichen928/SparseFusion.

Training with sparse annotations is known to reduce the performance of object detectors. Previous methods have focused on proxies for missing ground truth annotations in the form of pseudo-labels for unlabeled boxes. We observe that existing methods suffer at higher levels of sparsity in the data due to noisy pseudo-labels. To prevent this, we propose an end-to-end system that learns to separate the proposals into labeled and unlabeled regions using Pseudo-positive mining. While the labeled regions are processed as usual, self-supervised learning is used to process the unlabeled regions thereby preventing the negative effects of noisy pseudo-labels. This novel approach has multiple advantages such as improved robustness to higher sparsity when compared to existing methods. We conduct exhaustive experiments on five splits on the PASCAL-VOC and COCO datasets achieving state-of-the-art performance. We also unify various splits used across literature for this task and present a standardized benchmark. On average, we improve by 2.6, 3.9 and 9.6 mAP over previous state-of-the-art methods on three splits of increasing sparsity on COCO. Our project is publicly available at cs.umd.edu/ sakshams/SparseDet.

Camera-based 3D object detection in BEV (Bird's Eye View) space has drawn great attention over the past few years. Dense detectors typically follow a two-stage pipeline by first constructing a dense BEV feature and then performing object detection in BEV space, which suffers from complex view transformations and high computation cost. On the other side, sparse detectors follow a query-based paradigm without explicit dense BEV feature construction, but achieve worse performance than the dense counterparts. In this paper, we find that the key to mitigate this performance gap is the adaptability of the detector in both BEV and image space. To achieve this goal, we propose SparseBEV, a fully sparse 3D object detector that outperforms the dense counterparts. SparseBEV contains three key designs, which are (1) scale-adaptive self attention to aggregate features with adaptive receptive field in BEV space, (2) adaptive spatio-temporal sampling to generate sampling locations under the guidance of queries, and (3) adaptive mixing to decode the sampled features with dynamic weights from the queries. On the test split of nuScenes, SparseBEV achieves the state-of-the-art performance of 67.5 NDS. On the val split, SparseBEV achieves 55.8 NDS while maintaining a real-time inference speed of 23.5 FPS. Code is available at https://github.com/MCG-NJU/SparseBEV.

In the real world, image degradations caused by rain often exhibit a combination of rain streaks and raindrops, thereby increasing the challenges of recovering the underlying clean image. Note that the rain streaks and raindrops have diverse shapes, sizes, and locations in the captured image, and thus modeling the correlation relationship between irregular degradations caused by rain artifacts is a necessary prerequisite for image deraining. This paper aims to present an efficient and flexible mechanism to learn and model degradation relationships in a global view, thereby achieving a unified removal of intricate rain scenes. To do so, we propose a Sparse Sampling Transformer based on Uncertainty-Driven Ranking, dubbed UDR-S2Former. Compared to previous methods, our UDR-S2Former has three merits. First, it can adaptively sample relevant image degradation information to model underlying degradation relationships. Second, explicit application of the uncertainty-driven ranking strategy can facilitate the network to attend to degradation features and understand the reconstruction process. Finally, experimental results show that our UDR-S2Former clearly outperforms state-of-the-art methods for all benchmarks.

3D lane detection usually builds a dense correspondence between the front-view space and the BEV space to estimate lane points in the 3D space. 3D lanes only occupy a small ratio of the dense correspondence, while most correspondence belongs to the redundant background. This sparsity phenomenon bottlenecks valuable computation and raises the computation cost of building a high-resolution correspondence for accurate results. In this paper, we propose a sparse point-guided 3D lane detection, focusing on points related to 3D lanes. Our method runs in a coarse-to-fine manner, including coarse-level lane detection and iterative fine-level sparse point refinements. In coarse-level lane detection, we build a dense but efficient correspondence between the front view and BEV space at a very low resolution to compute coarse lanes. Then in fine-level sparse point refinement, we sample sparse points around coarse lanes to extract local features from the high-resolution front-view feature map. The high-resolution local information brought by sparse points refines 3D lanes in the BEV space hierarchically from low resolution to high resolution. The sparse point guides a more effective information flow and greatly promotes the SOTA result by 3 points on the overall F1-score and 6 points on several hard situations while reducing almost half memory cost and speeding up 2 times.

Pedestrian trajectory prediction is critical in many vision tasks but challenging due to the multimodality of the future trajectory. Most existing methods predict multimodal trajectories conditioned by goals (future endpoints) or instances (all future points). However, goal-conditioned methods ignore the intermediate process and instance-conditioned methods ignore the stochasticity of pedestrian motions. In this paper, we propose a simple yet effective Sparse Instance Conditioned Network (SICNet), which gives a balanced solution between goal-conditioned and instance-conditioned methods. Specifically, SICNet learns comprehensive sparse instances, i.e., representative points of the future trajectory, through a mask generated by a long short-term memory encoder and uses the memory mechanism to store and retrieve such sparse instances. Hence SICNet can decode the observed trajectory into the future prediction conditioned on the stored sparse instance. Moreover, we design a memory refinement module that refines the retrieved sparse instances from the memory to reduce memory recall errors. Extensive experiments on ETH-UCY and SDD datasets show that our method outperforms existing state-of-the-art methods. In addition, ablation studies demonstrate the superiority of our method compared with goal-conditioned and instance-conditioned approaches.

We propose an algorithm, 4DRegSDF, for the spacetime surface regularization to improve the fidelity of neural rendering and reconstruction in dynamic scenes. The key idea is to impose local rigidity on the deformable Signed Distance Function (SDF) for temporal coherency. Our approach works by (1) sampling points on the deformed surface by taking gradient steps toward the steepest direction along SDF, (2) extracting differential surface geometry, such as tangent plane or curvature, at each sample, and (3) adjusting the local rigidity at different timestamps. This enables our dynamic surface regularization to align 4D spacetime geometry via 3D canonical space more accurately. Experiments demonstrate that our 4DRegSDF achieves state-of-the-art performance in both reconstruction and rendering quality over synthetic and real-world datasets.

The combination of Neural Architecture Search (NAS) and quantization has proven successful in automatically designing low-FLOPs INT8 quantized neural networks (QNN). However, directly applying NAS to design accurate QNN models that achieve low latency on real-world devices leads to inferior performance. In this work, we identify that the poor INT8 latency is due to the quantization-unfriendly issue: the operator and configuration (e.g., channel width) choices in prior art search spaces lead to diverse quantization efficiency and can slow down the INT8 inference speed. To address this challenge, we propose SpaceEvo, an automatic method for designing a dedicated, quantization-friendly search space for each target hardware. The key idea of SpaceEvo is to automatically search hardware-preferred operators and configurations to construct the search space, guided by a metric called Q-T score to quantify how quantization-friendly a candidate search space is. We further train a quantized-for-all supernet over our discovered search space, enabling the searched models to be directly deployed without extra retraining or quantization. Our discovered models, SEQnet, establish new SOTA INT8 quantized accuracy under various latency constraints, achieving up to 10.1% accuracy improvement on ImageNet than prior art CNNs under the same latency. Extensive experiments on real devices show that SpaceEvo consistently outperforms manually-designed search spaces with up to 2.5x faster speed while achieving the same accuracy.

Recently, prompt-based learning has made impressive progress on image class-incremental learning, but it still lacks sufficient exploration in the video domain. In this paper, we will fill this gap by learning multiple prompts based on a powerful image-language pre-trained model, i.e., CLIP, making it fit for video class-incremental learning (VCIL). For this purpose, we present a space-time prompting approach (ST-Prompt) which contains two kinds of prompts, i.e., task-specific prompts and task-agnostic prompts. The task-specific prompts are to address the catastrophic forgetting problem by learning multi-grained prompts, i.e., spatial prompts, temporal prompts and comprehensive prompts, for accurate task identification. The task-agnostic prompts maintain a globally-shared prompt pool, which can empower the pre-trained image models with temporal perception abilities by exchanging contexts between frames. By this means, ST-Prompt can transfer the plentiful knowledge in the image-language pre-trained models to the VCIL task with only a tiny set of prompts to be optimized. To evaluate ST-Prompt, we conduct extensive experiments on three standard benchmarks. The results show that ST-Prompt can significantly surpass the state-of-the-art VCIL methods, especially it gains 9.06% on HMDB51 dataset under the 1*25 stage setting.