ICCV 2023 Papers

The performance of text-video retrieval has been significantly improved by vision-language cross-modal learning schemes. The typical solution is to directly align the global video-level and sentence-level features during learning, which would ignore the intrinsic video-text relations, i.e., a text description only corresponds to a spatio-temporal part of videos. Hence, the matching process should consider both fine-grained spatial content and various temporal semantic events. To this end, we propose a text-video learning framework with progressive spatio-temporal prototype matching. Specifically, the vanilla matching process is decomposed into two complementary phases: object-phrase prototype matching and event-sentence prototype matching. In the object-phrase prototype matching phase, a spatial prototype generation mechanism is developed to predict key patches or words, which are sparsely integrated into object or phrase prototypes. Importantly, optimizing the local alignment between object-phrase prototypes helps the model perceive spatial details. In the event-sentence prototype matching phase, we design a temporal prototype generation mechanism to associate intra-frame objects and interact inter-frame temporal relations. Such progressively generated event prototypes can reveal semantic diversity in videos for dynamic matching. Validated by comprehensive experiments, our method consistently outperforms the state-of-the-art methods on four video retrieval benchmarks.

Automatic perception of human behaviors during social interactions is crucial for AR/VR applications, and an essential component is estimation of plausible 3D human pose and shape of our social partners from the egocentric view. One of the biggest challenges of this task is severe body truncation due to close social distances in egocentric scenarios, which brings large pose ambiguities for unseen body parts. To tackle this challenge, we propose a novel scene-conditioned diffusion method to model the body pose distribution. Conditioned on the 3D scene geometry, the diffusion model generates bodies in plausible human-scene interactions, with the sampling guided by a physics-based collision score to further resolve human-scene interpenetrations. The classifier-free training enables flexible sampling with different conditions and enhanced diversity. A visibility-aware graph convolution model guided by per-joint visibility serves as the diffusion denoiser to incorporate inter-joint dependencies and per-body-part control. Extensive evaluations show that our method generates bodies in plausible interactions with 3D scenes, achieving both superior accuracy for visible joints and diversity for invisible body parts. The code is available at https://sanweiliti.github.io/egohmr/egohmr.html.

In this study, we focus on Continual Semantic Segmentation (CSS) and present a novel approach to tackle the issue of existing methods struggling to learn new classes. The primary challenge of CSS is to learn new knowledge while retaining old knowledge, which is commonly known as the rigidity-plasticity dilemma. Existing approaches strive to address this by carefully balancing the learning of new and old classes during training on new data. Differently, this work aims to avoid this dilemma fundamentally rather than handling the difficulties involved in it. Specifically, we reveal that this dilemma mainly arises from the greater fluctuation of knowledge for new classes because they have never been learned before the current step. Additionally, the data available in incremental steps are usually inadequate, which can impede the model's ability to learn discriminative features for both new and old classes. To address these challenges, we introduce a novel concept of pre-learning for future knowledge. Our approach entails optimizing the feature space and output space for unlabeled data, which thus enables the model to acquire knowledge for future classes. With this approach, updating the model for new classes becomes as smooth as for old classes, effectively avoiding the rigidity-plasticity dilemma. We conducted extensive experiments and the results demonstrate a significant improvement in the learning of new classes compared to previous state-of-the-art methods.

We introduce PointOdyssey, a large-scale synthetic dataset, and data generation framework, for the training and evaluation of long-term fine-grained tracking algorithms. Our goal is to advance the state-of-the-art by placing emphasis on long videos with naturalistic motion. Toward the goal of naturalism, we animate deformable characters using real-world motion capture data, we build 3D scenes to match the motion capture environments, and we render camera viewpoints using trajectories mined via structure-from-motion on real videos. We create combinatorial diversity by randomizing character appearance, motion profiles, materials, lighting, 3D assets, and atmospheric effects. Our dataset currently includes 104 videos, averaging 2,000 frames long, with orders of magnitude more correspondence annotations than prior work. We show that existing methods can be trained from scratch in our dataset and outperform the published variants. Finally, we introduce modifications to the PIPs point tracking method, greatly widening its temporal receptive field, which improves its performance on PointOdyssey as well as on two real-world benchmarks. Our data and code are publicly available at: https://pointodyssey.com.

Denoising diffusion models hold great promise for generating diverse and realistic human motions. However, existing motion diffusion models largely disregard the laws of physics in the diffusion process and often generate physically-implausible motions with pronounced artifacts such as floating, foot sliding, and ground penetration. This seriously impacts the quality of generated motions and limits their real-world application. To address this issue, we present a novel physics-guided motion diffusion model (PhysDiff), which incorporates physical constraints into the diffusion process. Specifically, we propose a physics-based motion projection module that uses motion imitation in a physics simulator to project the denoised motion of a diffusion step to a physically-plausible motion. The projected motion is further used in the next diffusion step to guide the denoising diffusion process. Intuitively, the use of physics in our model iteratively pulls the motion toward a physically-plausible space, which cannot be achieved by simple post-processing. Experiments on large-scale human motion datasets show that our approach achieves state-of-the-art motion quality and improves physical plausibility drastically (>78% for all datasets).

We consider the problem of imaging a dynamic scene over an extreme range of timescales simultaneously--seconds to picoseconds--and doing so passively, without much light, and without any timing signals from the light source(s) emitting it. Because existing flux estimation techniques for single-photon cameras break down in this regime, we develop a flux probing theory that draws insights from stochastic calculus to enable reconstruction of a pixel's time-varying flux from a stream of monotonically-increasing photon detection timestamps. We use this theory to (1) show that passive free-running SPAD cameras have an attainable frequency bandwidth that spans the entire DC-to-31 GHz range in low-flux conditions, (2) derive a novel Fourier-domain flux reconstruction algorithm that scans this range for frequencies with statistically-significant support in the timestamp data, and (3) ensure the algorithm's noise model remains valid even for very low photon counts or non-negligible dead times. We show the potential of this asynchronous imaging regime by experimentally demonstrating several never-seen-before abilities: (1) imaging a scene illuminated simultaneously by sources operating at vastly different speeds without synchronization (bulbs, projectors, multiple pulsed lasers), (2) passive non-line-of-sight video acquisition, and (3) recording ultra-wideband video, which can be played back later at 30 Hz to show everyday motions--but can also be played a billion times slower to show the propagation of light itself.

Given monocular videos, we build 3D models of articulated objects and environments whose 3D configurations satisfy dynamics and contact constraints. At its core, our method leverages differentiable physics simulation to aid visual reconstructions. We couple differentiable physics simulation with differentiable rendering via coordinate descent, which enables end-to-end optimization of, not only 3D reconstructions, but also physical system parameters from videos. We demonstrate the effectiveness of physics-informed reconstruction on monocular videos of quadruped animals and humans. It reduces reconstruction artifacts (e.g., scale ambiguity, unbalanced poses, and foot swapping) that are challenging to address by visual cues alone, and produces better foot contact estimation.

Personalized federated learning has received an upsurge of attention due to the mediocre performance of conventional federated learning (FL) over heterogeneous data. Unlike conventional FL which trains a single global consensus model, personalized FL allows different models for different clients. However, existing personalized FL algorithms only implicitly transfer the collaborative knowledge across the federation by embedding the knowledge into the aggregated model or regularization. We observed that this implicit knowledge transfer fails to maximize the potential of each client's empirical risk toward other clients. Based on our observation, in this work, we propose Personalized Global Federated Learning (PGFed), a novel personalized FL framework that enables each client to personalize its own global objective by explicitly and adaptively aggregating the empirical risks of itself and other clients. To avoid massive (O(N^2)) communication overhead and potential privacy leakage while achieving this, each client's risk is estimated through a first-order approximation for other clients' adaptive risk aggregation. On top of PGFed, we develop a momentum upgrade, dubbed PGFedMo, to more efficiently utilize clients' empirical risks. Our extensive experiments on four datasets under different federated settings show consistent improvements of PGFed over previous state-of-the-art methods. The code is publicly available at https://github.com/ljaiverson/pgfed.

Affine correspondences have traditionally been used to improve feature matching over wide baselines. While recent work has successfully used affine correspondences to solve various relative camera pose estimation problems, less attention has been given to their use in absolute pose estimation. We introduce the first general solution to the problem of estimating the pose of a calibrated camera given a single observation of an oriented point and an affine correspondence. The advantage of our approach (P1AC) is that it requires only a single correspondence, in comparison to the traditional point-based approach (P3P), significantly reducing the combinatorics in robust estimation. P1AC provides a general solution that removes restrictive assumptions made in prior work and is applicable to large-scale image-based localization. We propose a minimal solution to the P1AC problem and evaluate our novel solver on synthetic data, showing its numerical stability and performance under various types of noise. On standard image-based localization benchmarks we show that P1AC achieves more accurate results than the widely used P3P algorithm. Code for our method is available at https://github.com/jonathanventura/P1AC/.

Transfer learning is beneficial by allowing the expressive features of models pretrained on large-scale datasets to be finetuned for the target task of smaller, more domain-specific datasets. However, there is a concern that these pretrained models may come with their own biases which would propagate into the finetuned model. In this work, we investigate bias when conceptualized as both spurious correlations between the target task and a sensitive attribute as well as underrepresentation of a particular group in the dataset. Under both notions of bias, we find that (1) models finetuned on top of pretrained models can indeed inherit their biases, but (2) this bias can be corrected for through relatively minor interventions to the finetuning dataset, and often with a negligible impact to performance. Our findings imply that careful curation of the finetuning dataset is important for reducing biases on a downstream task, and doing so can even compensate for bias in the pretrained model.

Large-scale multi-modal pre-training models such as CLIP and PaLI exhibit strong generalization on various visual domains and tasks. However, existing image classification benchmarks often evaluate recognition on a specific domain (e.g., outdoor images) or a specific task (e.g., classifying plant species), which falls short of evaluating whether pre-trained foundational models are universal visual recognizers. To address this, we formally present the task of Open-domain Visual Entity recognitioN (OVEN), where a model need to link an image onto a Wikipedia entity with respect to a text query. We construct OVEN by re-purposing 14 existing datasets with all labels grounded onto one single label space: Wikipedia entities. OVEN challenges models to select among six million possible Wikipedia entities, making it a general visual recognition benchmark with largest number of labels. Our study on state-of-the-art pre-trained models reveals large headroom in generalizing to the massive-scale label space. We show that a PaLI-based auto-regressive visual recognition model performs surprisingly well, even on Wikipedia entities that have never been seen during fine-tuning. We also find existing pre-trained models yield different unique strengths: while PaLI-based models obtains higher overall performance, CLIP-based models are better at recognizing tail entities.

This paper aims for high-performance offline LiDAR-based 3D object detection. We first observe that experienced human annotators annotate objects from a track-centric perspective. They first label objects in a track with clear shapes, and then leverage the temporal coherence to infer the annotations of obscure objects. Drawing inspiration from this, we propose a high-performance offline detector in a track-centric perspective instead of the conventional object-centric perspective. Our method features a bidirectional tracking module and a track-centric learning module. Such a design allows our detector to infer and refine a complete track once the object is detected at a certain moment. We refer to this characteristic as "onCe detecTed, neveR Lost" and name the proposed system CTRL. Extensive experiments demonstrate the remarkable performance of our method, surpassing the human-level annotating accuracy and outperforming the previous state-of-the-art methods in the highly competitive Waymo Open Dataset leaderboard without model ensemble. The code is available at https://github.com/tusen-ai/SST.

Generalizing deep learning models to unknown target domain distribution with low latency has motivated research into test-time training/adaptation (TTT/TTA). Existing approaches often focus on improving test-time training performance under well-curated target domain data. As figured out in this work, many state-of-the-art methods fail to maintain the performance when the target domain is contaminated with strong out-of-distribution (OOD) data, a.k.a. open-world test-time training (OWTTT). The failure is mainly due to the inability to distinguish strong OOD samples from regular weak OOD samples. To improve the robustness of OWTTT we first develop an adaptive strong OOD pruning which improves the efficacy of the self-training TTT method. We further propose a way to dynamically expand the prototypes to represent strong OOD samples for an improved weak/strong OOD data separation. Finally, we regularize self-training with distribution alignment and the combination yields the state-of-the-art performance on 5 OWTTT benchmarks. The code is available at https://github.com/Yushu-Li/OWTTT.

Language-based object detection is a promising direction towards building a natural interface to describe objects in images that goes far beyond plain category names. While recent methods show great progress in that direction, proper evaluation is lacking. With OmniLabel, we propose a novel task definition, dataset, and evaluation metric. The task subsumes standard and open-vocabulary detection as well as referring expressions. With more than 30K unique object descriptions on over 25K images, OmniLabel provides a challenge benchmark with diverse and complex object descriptions in a naturally open-vocabulary setting. Moreover, a key differentiation to existing benchmarks is that our object descriptions can refer to one, multiple or even no object, hence, providing negative examples in free-form text. The proposed evaluation handles the large label space and judges performance via a modified average precision metric, which we validate by evaluating strong language-based baselines. OmniLabel indeed provides a challenging test bed for future research on language-based detection.

Domain Adaptive Object Detection (DAOD) transfers an object detector to a novel domain free of labels. However, in the real world, besides encountering novel scenes, novel domains always contain novel-class objects de facto, which are ignored in existing research. Thus, we formulate and study a more practical setting, Adaptive Open-set Object Detection (AOOD), considering both novel scenes and classes. Directly combing off-the-shelled cross-domain and open-set approaches is sub-optimal since their low-order dependence, e.g., the confidence score, is insufficient for the AOOD with two dimensions of novel information. To address this, we propose a novel Structured mOtif MAtching (SOMA) framework for AOOD, which models the high-order relation with motifs, i.e., a statistically significant subgraph, and formulates AOOD solution as motif matching to learn with high-order patterns. In a nutshell, SOMA consists of Structure-aware Novel-class Learning (SNL) and Structure-aware Transfer Learning (STL). As for SNL, we establish an instance-oriented graph to capture the class-independent object feature hidden in different base classes. Then, a high-order metric is proposed to match the most significant motif as high-order patterns, serving for motif-guided novel-class learning. In STL, we set up a semantic-oriented graph to model the class-dependent relation across domains, and match unlabelled objects with high-order motifs to align the cross-domain distribution with structural awareness. Extensive experiments demonstrate that the proposed SOMA achieves state-of-the-art performance. Codes will be released publicly for further study.

Generating realistic human 3D reconstructions using image or video data is essential for various communication and entertainment applications. While existing methods achieved impressive results for body and facial regions, realistic hair modeling still remains challenging due to its high mechanical complexity. This work proposes an approach capable of accurate hair geometry reconstruction at a strand level from a monocular video or multi-view images captured in uncontrolled lighting conditions. Our method has two stages, with the first stage performing joint reconstruction of coarse hair and bust shapes and hair orientation using implicit volumetric representations. The second stage then estimates a strand-level hair reconstruction by reconciling in a single optimization process the coarse volumetric constraints with hair strand and hairstyle priors learned from the synthetic data. To further increase the reconstruction fidelity, we incorporate image-based losses into the fitting process using a new differentiable renderer. The combined system, named Neural Haircut, achieves high realism and personalization of the reconstructed hairstyles.

We present a novel type of neural fields that uses general radial bases for signal representation. State-of-the-art neural fields typically rely on grid-based representations for storing local neural features and N-dimensional linear kernels for interpolating features at continuous query points. The spatial positions of their neural features are fixed on grid nodes and cannot well adapt to target signals. Our method instead builds upon general radial bases with flexible kernel position and shape, which have higher spatial adaptivity and can more closely fit target signals. To further improve the channel-wise capacity of radial basis functions, we propose to compose them with multi-frequency sinusoid functions. This technique extends a radial basis to multiple Fourier radial bases of different frequency bands without requiring extra parameters, facilitating the representation of details. Moreover, by marrying adaptive radial bases with grid-based ones, our hybrid combination inherits both adaptivity and interpolation smoothness. We carefully designed weighting schemes to let radial bases adapt to different types of signals effectively. Our experiments on 2D image and 3D signed distance field representation demonstrate the higher accuracy and compactness of our method than prior arts. When applied to neural radiance field reconstruction, our method achieves state-of-the-art rendering quality, with small model size and comparable training speed.

The long-standing theory that a colour-naming system evolves under dual pressure of efficient communication and perceptual mechanism is supported by more and more linguistic studies, including analysing four decades of diachronic data from the Nafaanra language. This inspires us to explore whether machine learning could evolve and discover a similar colour-naming system via optimising the communication efficiency represented by high-level recognition performance. Here, we propose a novel colour quantisation transformer, CQFormer, that quantises colour space while maintaining the accuracy of machine recognition on the quantised images. Given an RGB image, Annotation Branch maps it into an index map before generating the quantised image with a colour palette; meanwhile the Palette Branch utilises a key-point detection way to find proper colours in the palette among the whole colour space. By interacting with colour annotation, CQFormer is able to balance both the machine vision accuracy and colour perceptual structure such as distinct and stable colour distribution for discovered colour system. Very interestingly, we even observe the consistent evolution pattern between our artificial colour system and basic colour terms across human languages. Besides, our colour quantisation method also offers an efficient quantisation method that effectively compresses the image storage while maintaining high performance in high-level recognition tasks such as classification and detection. Extensive experiments demonstrate the superior performance of our method with extremely low bit-rate colours, showing potential to integrate into quantisation network to quantities from image to network activation. The source code is available at https://github.com/ryeocthiv/CQFormer.

Monocular depth estimation has drawn widespread attention from the vision community due to its broad applications. In this paper, we propose a novel physics (geometry)-driven deep learning framework for monocular depth estimation by assuming that 3D scenes are constituted by piece-wise planes. Particularly, we introduce a new normal-distance head that outputs pixel-level surface normal and plane-to-origin distance for deriving depth at each position. Meanwhile, the normal and distance are regularized by a developed plane-aware consistency constraint. We further integrate an additional depth head to improve the robustness of the proposed framework. To fully exploit the strengths of these two heads, we develop an effective contrastive iterative refinement module that refines depth in a complementary manner according to the depth uncertainty. Extensive experiments indicate that the proposed method exceeds previous state-of-the-art competitors on the NYU-Depth-v2, KITTI and SUN RGB-D datasets. Notably, it ranks 1st among all submissions on the KITTI depth prediction online benchmark at the submission time. The source code is available at https://github.com/ShuweiShao/NDDepth.