Shenghai Yuan

Lang Deng, Jianfei Yang, Shenghai Yuan et al. · berkeley

As an important biomarker for human identification, human gait can be collected at a distance by passive sensors without subject cooperation, which plays an essential role in crime prevention, security detection and other human identification applications. At present, most research works are based on cameras and computer vision techniques to perform gait recognition. However, vision-based methods are not reliable when confronting poor illuminations, leading to degrading performances. In this paper, we propose a novel multimodal gait recognition method, namely GaitFi, which leverages WiFi signals and videos for human identification. In GaitFi, Channel State Information (CSI) that reflects the multi-path propagation of WiFi is collected to capture human gaits, while videos are captured by cameras. To learn robust gait information, we propose a Lightweight Residual Convolution Network (LRCN) as the backbone network, and further propose the two-stream GaitFi by integrating WiFi and vision features for the gait retrieval task. The GaitFi is trained by the triplet loss and classification loss on different levels of features. Extensive experiments are conducted in the real world, which demonstrates that the GaitFi outperforms state-of-the-art gait recognition methods based on single WiFi or camera, achieving 94.2% for human identification tasks of 12 subjects.

Haozhi Cao, Yuecong Xu, Jianfei Yang et al.

Multi-modal unsupervised domain adaptation (MM-UDA) for 3D semantic segmentation is a practical solution to embed semantic understanding in autonomous systems without expensive point-wise annotations. While previous MM-UDA methods can achieve overall improvement, they suffer from significant class-imbalanced performance, restricting their adoption in real applications. This imbalanced performance is mainly caused by: 1) self-training with imbalanced data and 2) the lack of pixel-wise 2D supervision signals. In this work, we propose Multi-modal Prior Aided (MoPA) domain adaptation to improve the performance of rare objects. Specifically, we develop Valid Ground-based Insertion (VGI) to rectify the imbalance supervision signals by inserting prior rare objects collected from the wild while avoiding introducing artificial artifacts that lead to trivial solutions. Meanwhile, our SAM consistency loss leverages the 2D prior semantic masks from SAM as pixel-wise supervision signals to encourage consistent predictions for each object in the semantic mask. The knowledge learned from modal-specific prior is then shared across modalities to achieve better rare object segmentation. Extensive experiments show that our method achieves state-of-the-art performance on the challenging MM-UDA benchmark. Code will be available at https://github.com/AronCao49/MoPA.

Thien-Minh Nguyen, Yizhuo Yang, Tien-Dat Nguyen et al.

While UWB-based methods can achieve high localization accuracy in small-scale areas, their accuracy and reliability are significantly challenged in large-scale environments. In this paper, we propose a learning-based framework named ULOC for Ultra-Wideband (UWB) based localization in such complex large-scale environments. First, anchors are deployed in the environment without knowledge of their actual position. Then, UWB observations are collected when the vehicle travels in the environment. At the same time, map-consistent pose estimates are developed from registering (onboard self-localization) data with the prior map to provide the training labels. We then propose a network based on MAMBA that learns the ranging patterns of UWBs over a complex large-scale environment. The experiment demonstrates that our solution can ensure high localization accuracy on a large scale compared to the state-of-the-art. We release our source code to benefit the community at https://github.com/brytsknguyen/uloc.

Haozhi Cao, Yuecong Xu, Jianfei Yang et al.

Continual Test-Time Adaptation (CTTA) generalizes conventional Test-Time Adaptation (TTA) by assuming that the target domain is dynamic over time rather than stationary. In this paper, we explore Multi-Modal Continual Test-Time Adaptation (MM-CTTA) as a new extension of CTTA for 3D semantic segmentation. The key to MM-CTTA is to adaptively attend to the reliable modality while avoiding catastrophic forgetting during continual domain shifts, which is out of the capability of previous TTA or CTTA methods. To fulfill this gap, we propose an MM-CTTA method called Continual Cross-Modal Adaptive Clustering (CoMAC) that addresses this task from two perspectives. On one hand, we propose an adaptive dual-stage mechanism to generate reliable cross-modal predictions by attending to the reliable modality based on the class-wise feature-centroid distance in the latent space. On the other hand, to perform test-time adaptation without catastrophic forgetting, we design class-wise momentum queues that capture confident target features for adaptation while stochastically restoring pseudo-source features to revisit source knowledge. We further introduce two new benchmarks to facilitate the exploration of MM-CTTA in the future. Our experimental results show that our method achieves state-of-the-art performance on both benchmarks.

Tianchen Deng, Yue Pan, Shenghai Yuan et al.

In this paper, we provide a comprehensive overview of existing scene representation methods for robotics, covering traditional representations such as point clouds, voxels, signed distance functions (SDF), and scene graphs, as well as more recent neural representations like Neural Radiance Fields (NeRF), 3D Gaussian Splatting (3DGS), and the emerging Foundation Models. While current SLAM and localization systems predominantly rely on sparse representations like point clouds and voxels, dense scene representations are expected to play a critical role in downstream tasks such as navigation and obstacle avoidance. Moreover, neural representations such as NeRF, 3DGS, and foundation models are well-suited for integrating high-level semantic features and language-based priors, enabling more comprehensive 3D scene understanding and embodied intelligence. In this paper, we categorized the core modules of robotics into five parts (Perception, Mapping, Localization, Navigation, Manipulation). We start by presenting the standard formulation of different scene representation methods and comparing the advantages and disadvantages of scene representation across different modules. This survey is centered around the question: What is the best 3D scene representation for robotics? We then discuss the future development trends of 3D scene representations, with a particular focus on how the 3D Foundation Model could replace current methods as the unified solution for future robotic applications. The remaining challenges in fully realizing this model are also explored. We aim to offer a valuable resource for both newcomers and experienced researchers to explore the future of 3D scene representations and their application in robotics. We have published an open-source project on GitHub and will continue to add new works and technologies to this project.

Liuhan Chen, Zongjian Li, Bin Lin et al.

Variational Autoencoder (VAE), compressing videos into latent representations, is a crucial preceding component of Latent Video Diffusion Models (LVDMs). With the same reconstruction quality, the more sufficient the VAE's compression for videos is, the more efficient the LVDMs are. However, most LVDMs utilize 2D image VAE, whose compression for videos is only in the spatial dimension and often ignored in the temporal dimension. How to conduct temporal compression for videos in a VAE to obtain more concise latent representations while promising accurate reconstruction is seldom explored. To fill this gap, we propose an omni-dimension compression VAE, named OD-VAE, which can temporally and spatially compress videos. Although OD-VAE's more sufficient compression brings a great challenge to video reconstruction, it can still achieve high reconstructed accuracy by our fine design. To obtain a better trade-off between video reconstruction quality and compression speed, four variants of OD-VAE are introduced and analyzed. In addition, a novel tail initialization is designed to train OD-VAE more efficiently, and a novel inference strategy is proposed to enable OD-VAE to handle videos of arbitrary length with limited GPU memory. Comprehensive experiments on video reconstruction and LVDM-based video generation demonstrate the effectiveness and efficiency of our proposed methods.

Bin Lin, Yunyang Ge, Xinhua Cheng et al.

We introduce Open-Sora Plan, an open-source project that aims to contribute a large generation model for generating desired high-resolution videos with long durations based on various user inputs. Our project comprises multiple components for the entire video generation process, including a Wavelet-Flow Variational Autoencoder, a Joint Image-Video Skiparse Denoiser, and various condition controllers. Moreover, many assistant strategies for efficient training and inference are designed, and a multi-dimensional data curation pipeline is proposed for obtaining desired high-quality data. Benefiting from efficient thoughts, our Open-Sora Plan achieves impressive video generation results in both qualitative and quantitative evaluations. We hope our careful design and practical experience can inspire the video generation research community. All our codes and model weights are publicly available at \url{https://github.com/PKU-YuanGroup/Open-Sora-Plan}.

Shenghai Yuan, Jinfa Huang, Yujun Shi et al.

Recent advances in Text-to-Video generation (T2V) have achieved remarkable success in synthesizing high-quality general videos from textual descriptions. A largely overlooked problem in T2V is that existing models have not adequately encoded physical knowledge of the real world, thus generated videos tend to have limited motion and poor variations. In this paper, we propose \textbf{MagicTime}, a metamorphic time-lapse video generation model, which learns real-world physics knowledge from time-lapse videos and implements metamorphic generation. First, we design a MagicAdapter scheme to decouple spatial and temporal training, encode more physical knowledge from metamorphic videos, and transform pre-trained T2V models to generate metamorphic videos. Second, we introduce a Dynamic Frames Extraction strategy to adapt to metamorphic time-lapse videos, which have a wider variation range and cover dramatic object metamorphic processes, thus embodying more physical knowledge than general videos. Finally, we introduce a Magic Text-Encoder to improve the understanding of metamorphic video prompts. Furthermore, we create a time-lapse video-text dataset called \textbf{ChronoMagic}, specifically curated to unlock the metamorphic video generation ability. Extensive experiments demonstrate the superiority and effectiveness of MagicTime for generating high-quality and dynamic metamorphic videos, suggesting time-lapse video generation is a promising path toward building metamorphic simulators of the physical world. Code: https://github.com/PKU-YuanGroup/MagicTime

Shenghai Yuan, Yizhuo Yang, Thien Hoang Nguyen et al.

In response to the evolving challenges posed by small unmanned aerial vehicles (UAVs), which possess the potential to transport harmful payloads or independently cause damage, we introduce MMAUD: a comprehensive Multi-Modal Anti-UAV Dataset. MMAUD addresses a critical gap in contemporary threat detection methodologies by focusing on drone detection, UAV-type classification, and trajectory estimation. MMAUD stands out by combining diverse sensory inputs, including stereo vision, various Lidars, Radars, and audio arrays. It offers a unique overhead aerial detection vital for addressing real-world scenarios with higher fidelity than datasets captured on specific vantage points using thermal and RGB. Additionally, MMAUD provides accurate Leica-generated ground truth data, enhancing credibility and enabling confident refinement of algorithms and models, which has never been seen in other datasets. Most existing works do not disclose their datasets, making MMAUD an invaluable resource for developing accurate and efficient solutions. Our proposed modalities are cost-effective and highly adaptable, allowing users to experiment and implement new UAV threat detection tools. Our dataset closely simulates real-world scenarios by incorporating ambient heavy machinery sounds. This approach enhances the dataset's applicability, capturing the exact challenges faced during proximate vehicular operations. It is expected that MMAUD can play a pivotal role in advancing UAV threat detection, classification, trajectory estimation capabilities, and beyond. Our dataset, codes, and designs will be available in https://github.com/ntu-aris/MMAUD.

Zongjian Li, Bin Lin, Yang Ye et al.

Video Variational Autoencoder (VAE) encodes videos into a low-dimensional latent space, becoming a key component of most Latent Video Diffusion Models (LVDMs) to reduce model training costs. However, as the resolution and duration of generated videos increase, the encoding cost of Video VAEs becomes a limiting bottleneck in training LVDMs. Moreover, the block-wise inference method adopted by most LVDMs can lead to discontinuities of latent space when processing long-duration videos. The key to addressing the computational bottleneck lies in decomposing videos into distinct components and efficiently encoding the critical information. Wavelet transform can decompose videos into multiple frequency-domain components and improve the efficiency significantly, we thus propose Wavelet Flow VAE (WF-VAE), an autoencoder that leverages multi-level wavelet transform to facilitate low-frequency energy flow into latent representation. Furthermore, we introduce a method called Causal Cache, which maintains the integrity of latent space during block-wise inference. Compared to state-of-the-art video VAEs, WF-VAE demonstrates superior performance in both PSNR and LPIPS metrics, achieving 2x higher throughput and 4x lower memory consumption while maintaining competitive reconstruction quality. Our code and models are available at https://github.com/PKU-YuanGroup/WF-VAE.

Qiuhao Li, Shenghai Yuan

In the context of rapid advancements in industrial automation, vision-based robotic grasping plays an increasingly crucial role. In order to enhance visual recognition accuracy, the utilization of large-scale datasets is imperative for training models to acquire implicit knowledge related to the handling of various objects. Creating datasets from scratch is a time and labor-intensive process. Moreover, existing datasets often contain errors due to automated annotations aimed at expediency, making the improvement of these datasets a substantial research challenge. Consequently, several issues have been identified in the annotation of grasp bounding boxes within the popular Jacquard Grasp. We propose utilizing a Human-In-The-Loop(HIL) method to enhance dataset quality. This approach relies on backbone deep learning networks to predict object positions and orientations for robotic grasping. Predictions with Intersection over Union (IOU) values below 0.2 undergo an assessment by human operators. After their evaluation, the data is categorized into False Negatives(FN) and True Negatives(TN). FN are then subcategorized into either missing annotations or catastrophic labeling errors. Images lacking labels are augmented with valid grasp bounding box information, whereas images afflicted by catastrophic labeling errors are completely removed. The open-source tool Labelbee was employed for 53,026 iterations of HIL dataset enhancement, leading to the removal of 2,884 images and the incorporation of ground truth information for 30,292 images. The enhanced dataset, named the Jacquard V2 Grasping Dataset, served as the training data for a range of neural networks.

Yuzhi Lai, Shenghai Yuan, Youssef Nassar et al.

Effective Human-Robot Interaction (HRI) is crucial for future service robots in aging societies. Existing solutions are biased toward only well-trained objects, creating a gap when dealing with new objects. Currently, HRI systems using predefined gestures or language tokens for pretrained objects pose challenges for all individuals, especially elderly ones. These challenges include difficulties in recalling commands, memorizing hand gestures, and learning new names. This paper introduces NVP-HRI, an intuitive multi-modal HRI paradigm that combines voice commands and deictic posture. NVP-HRI utilizes the Segment Anything Model (SAM) to analyze visual cues and depth data, enabling precise structural object representation. Through a pre-trained SAM network, NVP-HRI allows interaction with new objects via zero-shot prediction, even without prior knowledge. NVP-HRI also integrates with a large language model (LLM) for multimodal commands, coordinating them with object selection and scene distribution in real time for collision-free trajectory solutions. We also regulate the action sequence with the essential control syntax to reduce LLM hallucination risks. The evaluation of diverse real-world tasks using a Universal Robot showcased up to 59.2\% efficiency improvement over traditional gesture control, as illustrated in the video https://youtu.be/EbC7al2wiAc. Our code and design will be openly available at https://github.com/laiyuzhi/NVP-HRI.git.

Hanfang Liang, Yizhuo Yang, Jinming Hu et al.

Compact UAV systems, while advancing delivery and surveillance, pose significant security challenges due to their small size, which hinders detection by traditional methods. This paper presents a cost-effective, unsupervised UAV detection method using spatial-temporal sequence processing to fuse multiple LiDAR scans for accurate UAV tracking in real-world scenarios. Our approach segments point clouds into foreground and background, analyzes spatial-temporal data, and employs a scoring mechanism to enhance detection accuracy. Tested on a public dataset, our solution placed 4th in the CVPR 2024 UG2+ Challenge, demonstrating its practical effectiveness. We plan to open-source all designs, code, and sample data for the research community github.com/lianghanfang/UnLiDAR-UAV-Est.

Zhenyuan Xiao, Yizhuo Yang, Guili Xu et al.

The increasing use of compact UAVs has created significant threats to public safety, while traditional drone detection systems are often bulky and costly. To address these challenges, we propose AV-DTEC, a lightweight self-supervised audio-visual fusion-based anti-UAV system. AV-DTEC is trained using self-supervised learning with labels generated by LiDAR, and it simultaneously learns audio and visual features through a parallel selective state-space model. With the learned features, a specially designed plug-and-play primary-auxiliary feature enhancement module integrates visual features into audio features for better robustness in cross-lighting conditions. To reduce reliance on auxiliary features and align modalities, we propose a teacher-student model that adaptively adjusts the weighting of visual features. AV-DTEC demonstrates exceptional accuracy and effectiveness in real-world multi-modality data. The code and trained models are publicly accessible on GitHub \url{https://github.com/AmazingDay1/AV-DETC}.

Tianchen Deng, Guole Shen, Xun Chen et al.

Neural implicit scene representations have recently shown promising results in dense visual SLAM. However, existing implicit SLAM algorithms are constrained to single-agent scenarios, and fall difficulties in large-scale scenes and long sequences. Existing NeRF-based multi-agent SLAM frameworks cannot meet the constraints of communication bandwidth. To this end, we propose the first distributed multi-agent collaborative neural SLAM framework with hybrid scene representation, distributed camera tracking, intra-to-inter loop closure, and online distillation for multiple submap fusion. A novel triplane-grid joint scene representation method is proposed to improve scene reconstruction. A novel intra-to-inter loop closure method is designed to achieve local (single-agent) and global (multi-agent) consistency. We also design a novel online distillation method to fuse the information of different submaps to achieve global consistency. Furthermore, to the best of our knowledge, there is no real-world dataset for NeRF-based/GS-based SLAM that provides both continuous-time trajectories groundtruth and high-accuracy 3D meshes groundtruth. To this end, we propose the first real-world Dense slam (DES) dataset covering both single-agent and multi-agent scenarios, ranging from small rooms to large-scale outdoor scenes, with high-accuracy ground truth for both 3D mesh and continuous-time camera trajectory. This dataset can advance the development of the research in both SLAM, 3D reconstruction, and visual foundation model. Experiments on various datasets demonstrate the superiority of the proposed method in both mapping, tracking, and communication. The dataset and code will open-source on https://github.com/dtc111111/mcnslam.

Bin Lin, Zongjian Li, Xinhua Cheng et al.

Although existing unified models achieve strong performance in vision-language understanding and text-to-image generation, they remain limited in addressing image perception and manipulation -- capabilities increasingly demanded in practical applications. Recently, OpenAI introduced the powerful GPT-4o-Image model, which showcases advanced capabilities in comprehensive image perception and manipulation, sparking widespread interest. Through carefully designed experiments, we observe that GPT-4o-Image likely relies on semantic encoders rather than VAEs for feature extraction, despite VAEs being commonly regarded as crucial for image manipulation tasks. Inspired by this insight, we propose UniWorld-V1, a unified generative framework built upon semantic features extracted from powerful multimodal large language models and contrastive semantic encoders. Using only 2.7M training data, UniWorld-V1 achieves impressive performance across diverse tasks, including image understanding, generation, manipulation, and perception. We fully open-source the UniWorld-V1 framework, including model weights, training and evaluation scripts, and datasets to promote reproducibility and further research.

Kun Cao, Muqing Cao, Shenghai Yuan et al.

This paper introduces a differential dynamic programming (DDP) based framework for polynomial trajectory generation for differentially flat systems. In particular, instead of using a linear equation with increasing size to represent multiple polynomial segments as in literature, we take a new perspective from state-space representation such that the linear equation reduces to a finite horizon control system with a fixed state dimension and the required continuity conditions for consecutive polynomials are automatically satisfied. Consequently, the constrained trajectory generation problem (both with and without time optimization) can be converted to a discrete-time finite-horizon optimal control problem with inequality constraints, which can be approached by a recently developed interior-point DDP (IPDDP) algorithm. Furthermore, for unconstrained trajectory generation with preallocated time, we show that this problem is indeed a linear-quadratic tracking (LQT) problem (DDP algorithm with exact one iteration). All these algorithms enjoy linear complexity with respect to the number of segments. Both numerical comparisons with state-of-the-art methods and physical experiments are presented to verify and validate the effectiveness of our theoretical findings. The implementation code will be open-sourced,

Thien-Minh Nguyen, Shenghai Yuan, Thien Hoang Nguyen et al.

Perception plays a crucial role in various robot applications. However, existing well-annotated datasets are biased towards autonomous driving scenarios, while unlabelled SLAM datasets are quickly over-fitted, and often lack environment and domain variations. To expand the frontier of these fields, we introduce a comprehensive dataset named MCD (Multi-Campus Dataset), featuring a wide range of sensing modalities, high-accuracy ground truth, and diverse challenging environments across three Eurasian university campuses. MCD comprises both CCS (Classical Cylindrical Spinning) and NRE (Non-Repetitive Epicyclic) lidars, high-quality IMUs (Inertial Measurement Units), cameras, and UWB (Ultra-WideBand) sensors. Furthermore, in a pioneering effort, we introduce semantic annotations of 29 classes over 59k sparse NRE lidar scans across three domains, thus providing a novel challenge to existing semantic segmentation research upon this largely unexplored lidar modality. Finally, we propose, for the first time to the best of our knowledge, continuous-time ground truth based on optimization-based registration of lidar-inertial data on large survey-grade prior maps, which are also publicly released, each several times the size of existing ones. We conduct a rigorous evaluation of numerous state-of-the-art algorithms on MCD, report their performance, and highlight the challenges awaiting solutions from the research community.

Tianchen Deng, Yaohui Chen, Leyan Zhang et al.

Recent work has shown that 3D Gaussian-based SLAM enables high-quality reconstruction, accurate pose estimation, and real-time rendering of scenes. However, these approaches are built on a tremendous number of redundant 3D Gaussian ellipsoids, leading to high memory and storage costs, and slow training speed. To address the limitation, we propose a compact 3D Gaussian Splatting SLAM system that reduces the number and the parameter size of Gaussian ellipsoids. A sliding window-based masking strategy is first proposed to reduce the redundant ellipsoids. Then we observe that the covariance matrix (geometry) of most 3D Gaussian ellipsoids are extremely similar, which motivates a novel geometry codebook to compress 3D Gaussian geometric attributes, i.e., the parameters. Robust and accurate pose estimation is achieved by a global bundle adjustment method with reprojection loss. Extensive experiments demonstrate that our method achieves faster training and rendering speed while maintaining the state-of-the-art (SOTA) quality of the scene representation.

Tianchen Deng, Nailin Wang, Chongdi Wang et al.

Dense scene reconstruction for photo-realistic view synthesis has various applications, such as VR/AR, autonomous vehicles. However, most existing methods have difficulties in large-scale scenes due to three core challenges: \textit{(a) inaccurate depth input.} Accurate depth input is impossible to get in real-world large-scale scenes. \textit{(b) inaccurate pose estimation.} Most existing approaches rely on accurate pre-estimated camera poses. \textit{(c) insufficient scene representation capability.} A single global radiance field lacks the capacity to effectively scale to large-scale scenes. To this end, we propose an incremental joint learning framework, which can achieve accurate depth, pose estimation, and large-scale scene reconstruction. A vision transformer-based network is adopted as the backbone to enhance performance in scale information estimation. For pose estimation, a feature-metric bundle adjustment (FBA) method is designed for accurate and robust camera tracking in large-scale scenes. In terms of implicit scene representation, we propose an incremental scene representation method to construct the entire large-scale scene as multiple local radiance fields to enhance the scalability of 3D scene representation. Extended experiments have been conducted to demonstrate the effectiveness and accuracy of our method in depth estimation, pose estimation, and large-scale scene reconstruction.

Jianping Li, Thien-Minh Nguyen, Shenghai Yuan et al.

Accurate and consistent construction of point clouds from LiDAR scanning data is fundamental for 3D modeling applications. Current solutions, such as multiview point cloud registration and LiDAR bundle adjustment, predominantly depend on the local plane assumption, which may be inadequate in complex environments lacking of planar geometries or substantial initial pose errors. To mitigate this problem, this paper presents a LiDAR bundle adjustment with progressive spatial smoothing, which is suitable for complex environments and exhibits improved convergence capabilities. The proposed method consists of a spatial smoothing module and a pose adjustment module, which combines the benefits of local consistency and global accuracy. With the spatial smoothing module, we can obtain robust and rich surface constraints employing smoothing kernels across various scales. Then the pose adjustment module corrects all poses utilizing the novel surface constraints. Ultimately, the proposed method simultaneously achieves fine poses and parametric surfaces that can be directly employed for high-quality point cloud reconstruction. The effectiveness and robustness of our proposed approach have been validated on both simulation and real-world datasets. The experimental results demonstrate that the proposed method outperforms the existing methods and achieves better accuracy in complex environments with low planar structures.

Siyu Chen, Kangcheng Liu, Chen Wang et al.

Visual Odometry (VO) is vital for the navigation of autonomous systems, providing accurate position and orientation estimates at reasonable costs. While traditional VO methods excel in some conditions, they struggle with challenges like variable lighting and motion blur. Deep learning-based VO, though more adaptable, can face generalization problems in new environments. Addressing these drawbacks, this paper presents a novel hybrid visual odometry (VO) framework that leverages pose-only supervision, offering a balanced solution between robustness and the need for extensive labeling. We propose two cost-effective and innovative designs: a self-supervised homographic pre-training for enhancing optical flow learning from pose-only labels and a random patch-based salient point detection strategy for more accurate optical flow patch extraction. These designs eliminate the need for dense optical flow labels for training and significantly improve the generalization capability of the system in diverse and challenging environments. Our pose-only supervised method achieves competitive performance on standard datasets and greater robustness and generalization ability in extreme and unseen scenarios, even compared to dense optical flow-supervised state-of-the-art methods.

Yufan Deng, Xun Guo, Yizhi Wang et al. · bytedance

Video generation has witnessed remarkable progress with the advent of deep generative models, particularly diffusion models. While existing methods excel in generating high-quality videos from text prompts or single images, personalized multi-subject video generation remains a largely unexplored challenge. This task involves synthesizing videos that incorporate multiple distinct subjects, each defined by separate reference images, while ensuring temporal and spatial consistency. Current approaches primarily rely on mapping subject images to keywords in text prompts, which introduces ambiguity and limits their ability to model subject relationships effectively. In this paper, we propose CINEMA, a novel framework for coherent multi-subject video generation by leveraging Multimodal Large Language Model (MLLM). Our approach eliminates the need for explicit correspondences between subject images and text entities, mitigating ambiguity and reducing annotation effort. By leveraging MLLM to interpret subject relationships, our method facilitates scalability, enabling the use of large and diverse datasets for training. Furthermore, our framework can be conditioned on varying numbers of subjects, offering greater flexibility in personalized content creation. Through extensive evaluations, we demonstrate that our approach significantly improves subject consistency, and overall video coherence, paving the way for advanced applications in storytelling, interactive media, and personalized video generation.

Zongjian Li, Zheyuan Liu, Qihui Zhang et al.

Instruction-based image editing has achieved remarkable progress; however, models solely trained via supervised fine-tuning often overfit to annotated patterns, hindering their ability to explore and generalize beyond training distributions. To this end, we introduce Edit-R1, a novel post-training framework for instruction-based image editing based on policy optimization. Specifically, we utilize Diffusion Negative-aware Finetuning (DiffusionNFT), a likelihood-free policy optimization method consistent with the flow matching forward process, thereby enabling the use of higher-order samplers and more efficient training. Another key challenge here is the absence of a universal reward model, resulting from the diverse nature of editing instructions and tasks. To bridge this gap, we employ a Multimodal Large Language Model (MLLM) as a unified, training-free reward model, leveraging its output logits to provide fine-grained feedback. Furthermore, we carefully design a low-variance group filtering mechanism to reduce MLLM scoring noise and stabilize optimization. \texttt{UniWorld-V2}, trained with this framework, achieves \textbf{state-of-the-art} results on the ImgEdit and GEdit-Bench benchmarks, scoring 4.49 and 7.83, respectively. Crucially, our framework is model-agnostic, delivering substantial performance gains when applied to diverse base models like Qwen-Image-Edit and FLUX-Kontext, demonstrating its wide applicability. Code and models are publicly available to support further research.

Wentao Zhao, Yihe Niu, Yanbo Wang et al.

This work presents UNO, a unified monocular visual odometry framework that enables robust and adaptable pose estimation across diverse environments, platforms, and motion patterns. Unlike traditional methods that rely on deployment-specific tuning or predefined motion priors, our approach generalizes effectively across a wide range of real-world scenarios, including autonomous vehicles, aerial drones, mobile robots, and handheld devices. To this end, we introduce a Mixture-of-Experts strategy for local state estimation, with several specialized decoders that each handle a distinct class of ego-motion patterns. Moreover, we introduce a fully differentiable Gumbel-Softmax module that constructs a robust inter-frame correlation graph, selects the optimal expert decoder, and prunes erroneous estimates. These cues are then fed into a unified back-end that combines pre-trained, scale-independent depth priors with a lightweight bundling adjustment to enforce geometric consistency. We extensively evaluate our method on three major benchmark datasets: KITTI (outdoor/autonomous driving), EuRoC-MAV (indoor/aerial drones), and TUM-RGBD (indoor/handheld), demonstrating state-of-the-art performance.

Pengyu Yin, Shenghai Yuan, Haozhi Cao et al.

We present a one-shot LiDAR global localization algorithm featuring semantic disambiguation ability based on a lightweight tri-layered scene graph. While landmark semantic registration-based methods have shown promising performance improvements in global localization compared with geometric-only methods, landmarks can be repetitive and misleading for correspondence establishment. We propose to mitigate this problem by modeling semantic distributions with continuous functions learned from a population of Gaussian processes. Compared with discrete semantic labels, the continuous functions capture finer-grained geo-semantic information and also provide more detailed metric information for correspondence establishment. We insert this continuous function as the middle layer between the object layer and the metric-semantic layer, forming a tri-layered 3D scene graph, serving as a light-weight yet performant backend for one-shot localization. We term our global localization pipeline Outram-GSF (Gaussian semantic field) and conduct a wide range of experiments on publicly available data sets, validating the superior performance against the current state-of-the-art.

Yunyang Ge, Xinhua Cheng, Chengshu Zhao et al.

In Image-to-Video (I2V) generation, a video is created using an input image as the first-frame condition. Existing I2V methods concatenate the full information of the conditional image with noisy latents to achieve high fidelity. However, the denoisers in these methods tend to shortcut the conditional image, which is known as conditional image leakage, leading to performance degradation issues such as slow motion and color inconsistency. In this work, we further clarify that conditional image leakage leads to overfitting to in-domain data and decreases the performance in out-of-domain scenarios. Moreover, we introduce Fourier-Guided Latent Shifting I2V, named FlashI2V, to prevent conditional image leakage. Concretely, FlashI2V consists of: (1) Latent Shifting. We modify the source and target distributions of flow matching by subtracting the conditional image information from the noisy latents, thereby incorporating the condition implicitly. (2) Fourier Guidance. We use high-frequency magnitude features obtained by the Fourier Transform to accelerate convergence and enable the adjustment of detail levels in the generated video. Experimental results show that our method effectively overcomes conditional image leakage and achieves the best generalization and performance on out-of-domain data among various I2V paradigms. With only 1.3B parameters, FlashI2V achieves a dynamic degree score of 53.01 on Vbench-I2V, surpassing CogVideoX1.5-5B-I2V and Wan2.1-I2V-14B-480P. Project page: https://pku-yuangroup.github.io/FlashI2V/

Junlong Ke, Qiying Hu, Shenghai Yuan et al.

Modern signal processing (SP) pipelines, whether model-based or data-driven, often constrained by complex and fragmented workflow, rely heavily on expert knowledge and manual engineering, and struggle with adaptability and generalization under limited data. In contrast, Large Language Models (LLMs) offer strong reasoning capabilities, broad general-purpose knowledge, in-context learning, and cross-modal transfer abilities, positioning them as powerful tools for automating and generalizing SP workflows. Motivated by these potentials, we introduce SignalLLM, the first general-purpose LLM-based agent framework for general SP tasks. Unlike prior LLM-based SP approaches that are limited to narrow applications or tricky prompting, SignalLLM introduces a principled, modular architecture. It decomposes high-level SP goals into structured subtasks via in-context learning and domain-specific retrieval, followed by hierarchical planning through adaptive retrieval-augmented generation (RAG) and refinement; these subtasks are then executed through prompt-based reasoning, cross-modal reasoning, code synthesis, model invocation, or data-driven LLM-assisted modeling. Its generalizable design enables the flexible selection of problem solving strategies across different signal modalities, task types, and data conditions. We demonstrate the versatility and effectiveness of SignalLLM through five representative tasks in communication and sensing, such as radar target detection, human activity recognition, and text compression. Experimental results show superior performance over traditional and existing LLM-based methods, particularly in few-shot and zero-shot settings.

Rui Qian, Haozhi Cao, Tianchen Deng et al.

Monocular 3D Semantic Scene Completion (SSC) is a challenging yet promising task that aims to infer dense geometric and semantic descriptions of a scene from a single image. While recent object-centric paradigms significantly improve efficiency by leveraging flexible 3D Gaussian primitives, they still rely heavily on a large number of randomly initialized primitives, which inevitably leads to 1) inefficient primitive initialization and 2) outlier primitives that introduce erroneous artifacts. In this paper, we propose SplatSSC, a novel framework that resolves these limitations with a depth-guided initialization strategy and a principled Gaussian aggregator. Instead of random initialization, SplatSSC utilizes a dedicated depth branch composed of a Group-wise Multi-scale Fusion (GMF) module, which integrates multi-scale image and depth features to generate a sparse yet representative set of initial Gaussian primitives. To mitigate noise from outlier primitives, we develop the Decoupled Gaussian Aggregator (DGA), which enhances robustness by decomposing geometric and semantic predictions during the Gaussian-to-voxel splatting process. Complemented with a specialized Probability Scale Loss, our method achieves state-of-the-art performance on the Occ-ScanNet dataset, outperforming prior approaches by over 6.3% in IoU and 4.1% in mIoU, while reducing both latency and memory consumption by more than 9.3%. The code will be released upon acceptance.

Shenghai Yuan, Jinfa Huang, Yongqi Xu et al.

We propose a novel text-to-video (T2V) generation benchmark, ChronoMagic-Bench, to evaluate the temporal and metamorphic capabilities of the T2V models (e.g. Sora and Lumiere) in time-lapse video generation. In contrast to existing benchmarks that focus on visual quality and textual relevance of generated videos, ChronoMagic-Bench focuses on the model's ability to generate time-lapse videos with significant metamorphic amplitude and temporal coherence. The benchmark probes T2V models for their physics, biology, and chemistry capabilities, in a free-form text query. For these purposes, ChronoMagic-Bench introduces 1,649 prompts and real-world videos as references, categorized into four major types of time-lapse videos: biological, human-created, meteorological, and physical phenomena, which are further divided into 75 subcategories. This categorization comprehensively evaluates the model's capacity to handle diverse and complex transformations. To accurately align human preference with the benchmark, we introduce two new automatic metrics, MTScore and CHScore, to evaluate the videos' metamorphic attributes and temporal coherence. MTScore measures the metamorphic amplitude, reflecting the degree of change over time, while CHScore assesses the temporal coherence, ensuring the generated videos maintain logical progression and continuity. Based on ChronoMagic-Bench, we conduct comprehensive manual evaluations of ten representative T2V models, revealing their strengths and weaknesses across different categories of prompts, and providing a thorough evaluation framework that addresses current gaps in video generation research. Moreover, we create a large-scale ChronoMagic-Pro dataset, containing 460k high-quality pairs of 720p time-lapse videos and detailed captions ensuring high physical pertinence and large metamorphic amplitude. [Homepage](https://pku-yuangroup.github.io/ChronoMagic-Bench/).

Haozhi Cao, Yuecong Xu, Pengyu Yin et al.

Multi-modal test-time adaptation (MM-TTA) adapts models to an unlabeled target domain by leveraging the complementary multi-modal inputs in an online manner. While previous MM-TTA methods for 3D segmentation offer a promising solution by leveraging self-refinement per frame, they suffer from two major limitations: 1) unstable frame-wise predictions caused by temporal inconsistency, and 2) consistently incorrect predictions that violate the assumption of reliable modality guidance. To address these limitations, this work introduces a comprehensive two-fold framework. Firstly, building upon our previous work ReLiable Spatial-temporal Voxels (Latte), we propose Latte++ that better suppresses the unstable frame-wise predictions with more informative geometric correspondences. Instead of utilizing a universal sliding window, Latte++ employs multi-window aggregation to capture more reliable correspondences to better evaluate the local prediction consistency of different semantic categories. Secondly, to tackle the consistently incorrect predictions, we propose Interactive Test-Time Adaptation (ITTA), a flexible add-on to empower effortless human feedback with existing MM-TTA methods. ITTA introduces a novel human-in-the-loop approach that efficiently integrates minimal human feedback through interactive segmentation, requiring only simple point clicks and bounding box annotations. Instead of using independent interactive networks, ITTA employs a lightweight promptable branch with a momentum gradient module to capture and reuse knowledge from scarce human feedback during online inference. Extensive experiments across five MM-TTA benchmarks demonstrate that ITTA achieves consistent and notable improvements with robust performance gains for target classes of interest in challenging imbalanced scenarios, while Latte++ provides complementary benefits for temporal stability.

Jianfei Yang, He Huang, Yunjiao Zhou et al.

4D human perception plays an essential role in a myriad of applications, such as home automation and metaverse avatar simulation. However, existing solutions which mainly rely on cameras and wearable devices are either privacy intrusive or inconvenient to use. To address these issues, wireless sensing has emerged as a promising alternative, leveraging LiDAR, mmWave radar, and WiFi signals for device-free human sensing. In this paper, we propose MM-Fi, the first multi-modal non-intrusive 4D human dataset with 27 daily or rehabilitation action categories, to bridge the gap between wireless sensing and high-level human perception tasks. MM-Fi consists of over 320k synchronized frames of five modalities from 40 human subjects. Various annotations are provided to support potential sensing tasks, e.g., human pose estimation and action recognition. Extensive experiments have been conducted to compare the sensing capacity of each or several modalities in terms of multiple tasks. We envision that MM-Fi can contribute to wireless sensing research with respect to action recognition, human pose estimation, multi-modal learning, cross-modal supervision, and interdisciplinary healthcare research.

Thien-Minh Nguyen, Shenghai Yuan, Muqing Cao et al.

In recent years, autonomous robots have become ubiquitous in research and daily life. Among many factors, public datasets play an important role in the progress of this field, as they waive the tall order of initial investment in hardware and manpower. However, for research on autonomous aerial systems, there appears to be a relative lack of public datasets on par with those used for autonomous driving and ground robots. Thus, to fill in this gap, we conduct a data collection exercise on an aerial platform equipped with an extensive and unique set of sensors: two 3D lidars, two hardware-synchronized global-shutter cameras, multiple Inertial Measurement Units (IMUs), and especially, multiple Ultra-wideband (UWB) ranging units. The comprehensive sensor suite resembles that of an autonomous driving car, but features distinct and challenging characteristics of aerial operations. We record multiple datasets in several challenging indoor and outdoor conditions. Calibration results and ground truth from a high-accuracy laser tracker are also included in each package. All resources can be accessed via our webpage https://ntu-aris.github.io/ntu_viral_dataset.

Thien-Minh Nguyen, Shenghai Yuan, Muqing Cao et al.

In this paper, we propose a tightly-coupled, multi-modal simultaneous localization and mapping (SLAM) framework, integrating an extensive set of sensors: IMU, cameras, multiple lidars, and Ultra-wideband (UWB) range measurements, hence referred to as VIRAL (visual-inertial-ranging-lidar) SLAM. To achieve such a comprehensive sensor fusion system, one has to tackle several challenges such as data synchronization, multi-threading programming, bundle adjustment (BA), and conflicting coordinate frames between UWB and the onboard sensors, so as to ensure real-time localization and smooth updates in the state estimates. To this end, we propose a two stage approach. In the first stage, lidar, camera, and IMU data on a local sliding window are processed in a core odometry thread. From this local graph, new key frames are evaluated for admission to a global map. Visual feature-based loop closure is also performed to supplement the global factor graph with loop constraints. When the global factor graph satisfies a condition on spatial diversity, the BA process will be triggered to update the coordinate transform between UWB and onboard SLAM systems. The system then seamlessly transitions to the second stage where all sensors are tightly integrated in the odometry thread. The capability of our system is demonstrated via several experiments on high-fidelity graphical-physical simulation and public datasets.

Yang Lyu, Muqing Cao, Shenghai Yuan et al.

Unmanned Aerial Vehicle (UAV) has already demonstrated its potential in many civilian applications, and the façade inspection is among the most promising ones. In this paper, we focus on enabling the autonomous perception and control of a small UAV for a façade inspection task. Specifically, we consider the perception as a planar object pose estimation problem by simplifying the building structure as concatenation of planes, and the control as an optimal reference tracking control problem. First, a vision based adaptive observer is proposed which can realize stable plane pose estimation under very mild observation conditions. Second, a model predictive controller is designed to achieve stable tracking and smooth transition in a multi-plane scenario, while the persistent excitation (PE) condition of the observer and the maneuver constraints of the UAV are satisfied. The proposed autonomous plane pose estimation and plane tracking methods are tested in both simulation and practical building fasçade inspection scenarios, which demonstrate their effectiveness and practicability.

Yang Lyu, Thien-Minh Nguyen, Liu Liu et al.

Although Simultaneous Localization and Mapping (SLAM) has been an active research topic for decades, current state-of-the-art methods still suffer from instability or inaccuracy due to feature insufficiency or its inherent estimation drift, in many civilian environments. To resolve these issues, we propose a navigation system combing the SLAM and prior-map-based localization. Specifically, we consider additional integration of line and plane features, which are ubiquitous and more structurally salient in civilian environments, into the SLAM to ensure feature sufficiency and localization robustness. More importantly, we incorporate general prior map information into the SLAM to restrain its drift and improve the accuracy. To avoid rigorous association between prior information and local observations, we parameterize the prior knowledge as low dimensional structural priors defined as relative distances/angles between different geometric primitives. The localization is formulated as a graph-based optimization problem that contains sliding-window-based variables and factors, including IMU, heterogeneous features, and structure priors. We also derive the analytical expressions of Jacobians of different factors to avoid the automatic differentiation overhead. To further alleviate the computation burden of incorporating structural prior factors, a selection mechanism is adopted based on the so-called information gain to incorporate only the most effective structure priors in the graph optimization. Finally, the proposed framework is extensively tested on synthetic data, public datasets, and, more importantly, on the real UAV flight data obtained from a building inspection task. The results show that the proposed scheme can effectively improve the accuracy and robustness of localization for autonomous robots in civilian applications.

Thien-Minh Nguyen, Muqing Cao, Shenghai Yuan et al.

In recent years, thanks to the continuously reduced cost and weight of 3D Lidar, the applications of this type of sensor in robotics community have become increasingly popular. Despite many progresses, estimation drift and tracking loss are still prevalent concerns associated with these systems. However, in theory these issues can be resolved with the use of some observations to fixed landmarks in the environments. This motivates us to investigate a tightly coupled sensor fusion scheme of Ultra-Wideband (UWB) range measurements with Lidar and inertia measurements. First, data from IMU, Lidar and UWB are associated with the robot's states on a sliding windows based on their timestamps. Then, we construct a cost function comprising of factors from UWB, Lidar and IMU preintegration measurements. Finally an optimization process is carried out to estimate the robot's position and orientation. Via some real world experiments, we show that the method can effectively resolve the drift issue, while only requiring two or three anchors deployed in the environment.

Thien-Minh Nguyen, Shenghai Yuan, Muqing Cao et al.

In recent years, Onboard Self Localization (OSL) methods based on cameras or Lidar have achieved many significant progresses. However, some issues such as estimation drift and feature-dependence still remain inherent limitations. On the other hand, infrastructure-based methods can generally overcome these issues, but at the expense of some installation cost. This poses an interesting problem of how to effectively combine these methods, so as to achieve localization with long-term consistency as well as flexibility compared to any single method. To this end, we propose a comprehensive optimization-based estimator for 15-dimensional state of an Unmanned Aerial Vehicle (UAV), fusing data from an extensive set of sensors: inertial measurement units (IMUs), Ultra-Wideband (UWB) ranging sensors, and multiple onboard Visual-Inertial and Lidar odometry subsystems. In essence, a sliding window is used to formulate a sequence of robot poses, where relative rotational and translational constraints between these poses are observed in the IMU preintegration and OSL observations, while orientation and position are coupled in body-offset UWB range observations. An optimization-based approach is developed to estimate the trajectory of the robot in this sliding window. We evaluate the performance of the proposed scheme in multiple scenarios, including experiments on public datasets, high-fidelity graphical-physical simulator, and field-collected data from UAV flight tests. The result demonstrates that our integrated localization method can effectively resolve the drift issue, while incurring minimal installation requirements.