ICCV 2023 论文列表

Human Pose Estimation (HPE) is widely used in various fields, including motion analysis, healthcare, and virtual reality. However, the great expenses of labeled real-world datasets present a significant challenge for HPE. To overcome this, one approach is to train HPE models on synthetic datasets and then perform domain adaptation (DA) on real-world data. Unfortunately, existing DA methods for HPE neglect data privacy and security by using both source and target data in the adaptation process. To this end, we propose a new task, named source-free domain adaptive HPE, which aims to address the challenges of cross-domain learning of HPE without access to source data during the adaptation process. We further propose a novel framework that consists of three models: source model, intermediate model, and target model, which explores the task from both source-protect and target-relevant perspectives. The source-protect module preserves source information more effectively while resisting noise, and the target-relevant module reduces the sparsity of spatial representations by building a novel spatial probability space, and pose-specific contrastive learning and information maximization are proposed on the basis of this space. Comprehensive experiments on several domain adaptive HPE benchmarks show that the proposed method outperforms existing approaches by a considerable margin. The codes are available at https://github.com/davidpengucf/SFDAHPE.

Depth cues are known to be useful for visual perception. However, direct measurement of depth is often impracticable. Fortunately, though, modern learning-based methods offer promising depth maps by inference in the wild. In this work, we adapt such depth inference models for object segmentation using the objects' "pop-out" prior in 3D. The "pop-out" is a simple composition prior that assumes objects reside on the background surface. Such compositional prior allows us to reason about objects in the 3D space. More specifically, we adapt the inferred depth maps such that objects can be localized using only 3D information. Such separation, however, requires knowledge about contact surface which we learn using the weak supervision of the segmentation mask. Our intermediate representation of contact surface, and thereby reasoning about objects purely in 3D, allows us to better transfer the depth knowledge into semantics. The proposed adaptation method uses only the depth model without needing the source data used for training, making the learning process efficient and practical. Our experiments on eight datasets of two challenging tasks, namely salient object detection and camouflaged object detection, consistently demonstrate the benefit of our method in terms of both performance and generalizability. The source code is publicly available at https://github.com/Zongwei97/PopNet.

Humans can easily perceive the direction of sound sources in a visual scene, termed sound source localization. Recent studies on learning-based sound source localization have mainly explored the problem from a localization perspective. However, prior arts and existing benchmarks do not account for a more important aspect of the problem, cross-modal semantic understanding, which is essential for genuine sound source localization. Cross-modal semantic understanding is important in understanding semantically mismatched audio-visual events, e.g., silent objects, or off-screen sounds. To account for this, we propose a cross-modal alignment task as a joint task with sound source localization to better learn the interaction between audio and visual modalities. Thereby, we achieve high localization performance with strong cross-modal semantic understanding. Our method outperforms the state-of-the-art approaches in both sound source localization and cross-modal retrieval. Our work suggests that jointly tackling both tasks is necessary to conquer genuine sound source localization.

The images and sounds that we perceive undergo subtle but geometrically consistent changes as we rotate our heads. In this paper, we use these cues to solve a problem we call Sound Localization from Motion (SLfM): jointly estimating camera rotation and localizing sound sources. We learn to solve these tasks solely through self-supervision. A visual model predicts camera rotation from a pair of images, while an audio model predicts the direction of sound sources from binaural sounds. We train these models to generate predictions that agree with one another. At test time, the models can be deployed independently. To obtain a feature representation that is well-suited to solving this challenging problem, we also propose a method for learning an audio-visual representation through cross-view binauralization: estimating binaural sound from one view, given images and sound from another. Our model can successfully estimate accurate rotations on both real and synthetic scenes, and localize sound sources with accuracy competitive with state-of-the-art self-supervised approaches.

We propose Social Diffusion, a novel method for short-term and long-term forecasting of the motion of multiple persons as well as their social interactions. Jointly forecasting motions for multiple persons involved in social activities is inherently a challenging problem due to the interdependencies between individuals. In this work, we leverage a diffusion model conditioned on motion histories and causal temporal convolutional networks to forecast individually and contextually plausible motions for all participants. The contextual plausibility is achieved via an order-invariant aggregation function. As a second contribution, we design a new evaluation protocol that measures the plausibility of social interactions which we evaluate on the Haggling dataset, which features a challenging social activity where people are actively taking turns to talk and switching their attention. We evaluate our approach on four datasets for multi-person forecasting where our approach outperforms the state-of-the-art in terms of motion realism and contextual plausibility.

Reinterpretable cameras are defined by their post-processing capabilities that exceed traditional imaging. We present "SoDaCam" that provides reinterpretable cameras at the granularity of photons, from photon-cubes acquired by single-photon devices. Photon-cubes represent the spatio-temporal detections of photons as a sequence of binary frames, at frame-rates as high as 100 kHz. We show that simple transformations of the photon-cube, or photon-cube projections, provide the functionality of numerous imaging systems including: exposure bracketing, flutter shutter cameras, video compressive systems, event cameras, and even cameras that move during exposure. Our photon-cube projections offer the flexibility of being software-defined constructs that are only limited by what is computable, and shot-noise. We exploit this flexibility to provide new capabilities for the emulated cameras. As an added benefit, our projections provide camera-dependent compression of photon-cubes, which we demonstrate using an implementation of our projections on a novel compute architecture that is designed for single-photon imaging.

Snowfall is a common weather phenomenon that can severely affect computer vision tasks by obscuring objects and scenes. However, existing deep learning-based snow removal methods are designed for single images only. In this paper, we target a more complex task -- video snow removal, which aims to restore the clear video from the snowy video. To facilitate this task, we propose the first high-quality video dataset, which simulates realistic physical characteristics of snow and haze using a rendering engine and augmentation techniques. We also develop a deep learning framework for video snow removal. It involves Specifically, we propose a snow-query temporal aggregation module and a snow-aware contrastive learning loss function. The module aggregates features between video frames and removes snow effectively, while the loss function helps identify and eliminate snow features. We conduct extensive experiments and demonstrate that our proposed dataset is more realistic than previous datasets, and the models trained on it achieve better performance in real-world snowing images. Our proposed method outperforms state-of-the-art video and image-based methods on both synthetic and real snowy videos.

The task of few-shot GAN adaptation aims to adapt a pre-trained GAN model to a small dataset with very few training images. While existing methods perform well when the dataset for pre-training is structurally similar to the target dataset, the approaches suffer from training instabilities or memorization issues when the objects in the two domains have a very different structure. To mitigate this limitation, we propose a new smoothness similarity regularization that transfers the inherently learned smoothness of the pre-trained GAN to the few-shot target domain even if the two domains are very different. We evaluate our approach by adapting an unconditional and a class-conditional GAN to diverse few-shot target domains. Our proposed method significantly outperforms prior few-shot GAN adaptation methods in the challenging case of structurally dissimilar source-target domains, while performing on par with the state of the art for similar source-target domains.

The past few years have witnessed the immense success of object detection, while current excellent detectors struggle on tackling size-limited instances. Concretely, the well-known challenge of low overlaps between the priors and object regions leads to a constrained sample pool for optimization, and the paucity of discriminative information further aggravates the recognition. To alleviate the aforementioned issues, we propose CFINet, a two-stage framework tailored for small object detection based on the Coarse-to-fine pipeline and Feature Imitation learning. Firstly, we introduce Coarse-to-fine RPN (CRPN) to ensure sufficient and high-quality proposals for small objects through the dynamic anchor selection strategy and cascade regression. Then, we equip the conventional detection head with a Feature Imitation (FI) branch to facilitate the region representations of size-limited instances that perplex the model in an imitation manner. Moreover, an auxiliary imitation loss following supervised contrastive learning paradigm is devised to optimize this branch. When integrated with Faster RCNN, CFINet achieves state-of-the-art performance on the large-scale small object detection benchmarks, SODA-D and SODA-A, underscoring its superiority over baseline detector and other mainstream detection approaches. Code is available at https://github.com/shaunyuan22/CFINet.

Although 3D perception for autonomous vehicles has focused on frontal-view information, more than half of fatal accidents occur due to side impacts in practice (e.g., T-bone crash). Motivated by this fact, we investigate the problem of side-view depth estimation, especially for monocular fisheye cameras, which provide wide FoV information. However, since fisheye cameras head road areas, it observes road areas mostly and results in severe distortion on object areas, such as vehicles or pedestrians. To alleviate these issues, we propose a new fisheye depth estimation network, SlaBins, that infers an accurate and dense depth map based on a geometric property of road environments; most objects are standing (i.e., orthogonal) on the road environments. Concretely, we introduce a slanted multi-cylindrical image (MCI) representation, which allows us to describe a distance as a radius to a cylindrical layer orthogonal to the ground regardless of the camera viewing direction. Based on the slanted MCI, we estimate a set of adaptive bins and a per-pixel probability map for depth estimation. Then by combining it with the estimated slanted angle of viewing direction, we directly infer a dense and accurate depth map for fisheye cameras. Experiments demonstrate that SlaBins outperforms the state-of-the-art methods in both qualitative and quantitative evaluation on the SynWoodScape and KITTI-360 depth datasets.

In this paper, we propose Skip-Plan, a condensed action space learning method for procedure planning in instructional videos. Current procedure planning methods all stick to the state-action pair prediction at every timestep and generate actions adjacently. Although it coincides with human intuition, such a methodology consistently struggles with high-dimensional state supervision and error accumulation on action sequences. In this work, we abstract the procedure planning problem as a mathematical chain model. By skipping uncertain nodes and edges in action chains, we transfer long and complex sequence functions into short but reliable ones in two ways. First, we skip all the intermediate state supervision and only focus on action predictions. Second, we decompose relatively long chains into multiple short sub-chains by skipping unreliable intermediate actions. By this means, our model explores all sorts of reliable sub-relations within an action sequence in the condensed action space. Extensive experiments show Skip-Plan achieves state-of-the-art performance on the CrossTask and COIN benchmarks for procedure planning.

We present Skill Transformer, an approach for solving long-horizon robotic tasks by combining conditional sequence modeling and skill modularity. Conditioned on egocentric and proprioceptive observations of a robot, Skill Transformer is trained end-to-end to predict both a high-level skill (e.g., navigation, picking, placing), and a whole-body low-level action (e.g., base and arm motion), using a transformer architecture and demonstration trajectories that solve the full task. It retains the composability and modularity of the overall task through a skill predictor module while reasoning about low-level actions and avoiding hand-off errors, common in modular approaches. We test Skill Transformer on an embodied rearrangement benchmark and find it performs robust task planning and low-level control in new scenarios, achieving a 2.5x higher success rate than baselines in hard rearrangement problems.

Diffusion probabilistic models have achieved remarkable success in text guided image generation. However, generating 3D shapes is still challenging due to the lack of sufficient data containing 3D models along with their descriptions. Moreover, text based descriptions of 3D shapes are inherently ambiguous and lack details. In this paper, we propose a sketch and text guided probabilistic diffusion model for colored point cloud generation that conditions the denoising process jointly with a hand drawn sketch of the object and its textual description. We incrementally diffuse the point coordinates and color values in a joint diffusion process to reach a Gaussian distribution. Colored point cloud generation thus amounts to learning the reverse diffusion process, conditioned by the sketch and text, to iteratively recover the desired shape and color. Specifically, to learn effective sketch-text embedding, our model adaptively aggregates the joint embedding of text prompt and the sketch based on a capsule attention network. Our model uses staged diffusion to generate the shape and then assign colors to different parts conditioned on the appearance prompt while preserving precise shapes from the first stage. This gives our model the flexibility to extend to multiple tasks, such as appearance re-editing and part segmentation. Experimental results demonstrate that our model outperforms the recent state-of-the-art in point cloud generation.

Skeleton sequence representation learning has shown great advantages for action recognition due to its promising ability to model human joints and topology. However, the current methods usually require sufficient labeled data for training computationally expensive models. Moreover, these methods ignore how to utilize the fine-grained dependencies among different skeleton joints to pre-train an efficient skeleton sequence learning model that can generalize well across different datasets. In this paper, we propose an efficient skeleton sequence learning framework, named Skeleton Sequence Learning (SSL). To comprehensively capture the human pose and obtain discriminative skeleton sequence representation, we build an asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE, which embeds skeleton joint sequence into graph convolutional network and reconstructs the masked skeleton joints and edges based on the prior human topology knowledge. Then, the pre-trained SkeletonMAE encoder is integrated with the Spatial-Temporal Representation Learning (STRL) module to build the SSL framework. Extensive experimental results show that our SSL generalizes well across different datasets and outperforms the state-of-the-art self-supervised skeleton-based methods on FineGym, Diving48, NTU 60 and NTU 120 datasets. Moreover, we obtain comparable performance to some fully supervised methods. The code is avaliable at https://github.com/HongYan1123/SkeletonMAE.

We present SkeleTR, a new framework for skeleton-based action recognition. In contrast to prior work, which focuses mainly on controlled environments, we target in-the-wild scenarios that typically involve a variable number of people and various forms of interaction between people. SkeleTR works with a two-stage paradigm. It first models the intra-person skeleton dynamics for each skeleton sequence with graph convolutions, and then uses stacked Transformer encoders to capture person interactions that are important for action recognition in the wild. To mitigate the negative impact of inaccurate skeleton associations, SkeleTR takes relative short skeleton sequences as input and increases the number of sequences. As a unified solution, SkeleTR can be directly applied to multiple skeleton-based action tasks, including video-level action classification, instance-level action detection, and group-level activity recognition. It also enables transfer learning and joint training across different action tasks and datasets, which results in performance improvement. When evaluated on various skeleton-based action recognition benchmarks, SkeleTR achieves the state-of-the-art performance.

Virtual try-on tasks aim at synthesizing realistic try-on results by trying target clothes on humans. Most previous works relied on the Thin Plate Spline or appearance flows to warp clothes to fit human body shapes. However, both approaches cannot handle complex warping, leading to over distortion or misalignment. Furthermore, there is a critical unaddressed challenge of adjusting clothing sizes for try-on. To tackle these issues, we propose a Clothing-Oriented Transformation Try-On Network (COTTON). COTTON leverages clothing structure with landmarks and segmentation to design a novel landmark-guided transformation for precisely deforming clothes, allowing for size adjustment during try-on. Additionally, to properly remove the clothing region from the human image without losing significant human characteristics, we propose a clothing elimination policy based on both transformed clothes and human segmentation. This method enables users to try on clothes tucked-in or untucked while retaining more human characteristics. Both qualitative and quantitative results show that COTTON outperforms the state-of-the-art high-resolution virtual try-on approaches. All the code is available at https://github.com/cotton6/COTTON-size-does-matter.

3D-aware image synthesis encompasses a variety of tasks, such as scene generation and novel view synthesis from images. Despite numerous task-specific methods, developing a comprehensive model remains challenging. In this paper, we present SSDNeRF, a unified approach that employs an expressive diffusion model to learn a generalizable prior of neural radiance fields (NeRF) from multi-view images of diverse objects. Previous studies have used two-stage approaches that rely on pretrained NeRFs as real data to train diffusion models. In contrast, we propose a new single-stage training paradigm with an end-to-end objective that jointly optimizes a NeRF auto-decoder and a latent diffusion model, enabling simultaneous 3D reconstruction and prior learning, even from sparsely available views. At test time, we can directly sample the diffusion prior for unconditional generation, or combine it with arbitrary observations of unseen objects for NeRF reconstruction. SSDNeRF demonstrates robust results comparable to or better than leading task-specific methods in unconditional generation and single/sparse-view 3D reconstruction.

The reflection superposition phenomenon is complex and widely distributed in the real world, which derives various simplified linear and nonlinear formulations of the problem. In this paper, based on the investigation of the weaknesses of existing models, we propose a more general form of the superposition model by introducing a learnable residue term, which can effectively capture residual information during decomposition, guiding the separated layers to be complete. In order to fully capitalize on its advantages, we further design the network structure elaborately, including a novel dual-stream interaction mechanism and a powerful decomposition network with a semantic pyramid encoder. Extensive experiments and ablation studies are conducted to verify our superiority over state-of-the-art approaches on multiple real-world benchmark datasets.

Single image defocus deblurring (SIDD) is a challenging task due to the spatially-varying nature of defocus blur, characterized by per-pixel point spread functions (PSFs). Existing deep-learning-based methods for SIDD are limited by either over-fitting due to the lack of model constraints or under-parametrization that restricts their applicability to real-world images. To address the limitations, this paper proposes an interpretable approach that explicitly predicts inverse kernels with structural regularization. Motivated by the observation that defocus PSFs within an image often have similar shapes but different sizes, we represent the inverse kernels linearly over a multi-scale dictionary parameterized by implicit neural representations. We predict the corresponding representation coefficients via a duplex scale-recurrent neural network that jointly performs fine-to-coarse and coarse-to-fine estimations. Extensive experiments demonstrate that our approach achieves excellent performance using a lightweight model.