Yunze Liu

Guanyi Qin, Jie Liang, Bingbing Zhang et al. · baidu

In this paper, we present an overview of the NTIRE 2026 challenge on the 3rd Restore Any Image Model in the Wild, specifically focusing on Track 1: Professional Image Quality Assessment. Conventional Image Quality Assessment (IQA) typically relies on scalar scores. By compressing complex visual characteristics into a single number, these methods fundamentally struggle to distinguish subtle differences among uniformly high-quality images. Furthermore, they fail to articulate why one image is superior, lacking the reasoning capabilities required to provide guidance for vision tasks. To bridge this gap, recent advancements in Multimodal Large Language Models (MLLMs) offer a promising paradigm. Inspired by this potential, our challenge establishes a novel benchmark exploring the ability of MLLMs to mimic human expert cognition in evaluating high-quality image pairs. Participants were tasked with overcoming critical bottlenecks in professional scenarios, centering on two primary objectives: (1) Comparative Quality Selection: reliably identifying the visually superior image within a high-quality pair; and (2) Interpretative Reasoning: generating grounded, expert-level explanations that detail the rationale behind the selection. In total, the challenge attracted nearly 200 registrations and over 2,500 submissions. The top-performing methods significantly advanced the state of the art in professional IQA. The challenge dataset is available at https://github.com/narthchin/RAIM-PIQA, and the official homepage is accessible at https://www.codabench.org/competitions/12789/.

Yunze Liu, Chi-Hao Wu, Enmin Zhou et al.

Unified multimodal embedding spaces have become the standard interface for cross-modal retrieval and multimodal RAG, and recent audio-video-text (AVT) encoders extend this setting to three modalities. Such encoders can produce a joint (T,V,A) embedding whenever all three modalities are available, but standard pairwise InfoNCE objectives leave this signal unused during training. We close this gap with fusion-as-teacher distillation, which treats a stop-gradient copy of the fused embedding as a teacher signal for the single-modal embeddings, paired with a Tuple-InfoNCE term that supervises the fused embedding directly. We instantiate this objective as OmniRetriever-7B. Across six zero-shot retrieval benchmarks, OmniRetriever-7B surpasses the closed-source Gemini Embedding 2 by 13.3-18.0 R@1 on Clotho and SoundDescs, and reaches the contemporary zero-shot specialist band of open video-text encoders on MSR-VTT and MSVD. To stress-test joint representations, we further release OmniRetriever-Bench, a 12-direction AVT retrieval benchmark totaling 3782 triples; on it OmniRetriever-7B attains AVG-all 34.84, improving over Gemini Embedding 2 by 1.72 and over the best prior open-source AVT method by 8.03.

Yunze Liu, Yun Liu, Che Jiang et al.

We present HOI4D, a large-scale 4D egocentric dataset with rich annotations, to catalyze the research of category-level human-object interaction. HOI4D consists of 2.4M RGB-D egocentric video frames over 4000 sequences collected by 4 participants interacting with 800 different object instances from 16 categories over 610 different indoor rooms. Frame-wise annotations for panoptic segmentation, motion segmentation, 3D hand pose, category-level object pose and hand action have also been provided, together with reconstructed object meshes and scene point clouds. With HOI4D, we establish three benchmarking tasks to promote category-level HOI from 4D visual signals including semantic segmentation of 4D dynamic point cloud sequences, category-level object pose tracking, and egocentric action segmentation with diverse interaction targets. In-depth analysis shows HOI4D poses great challenges to existing methods and produces great research opportunities.

Hao Wen, Yunze Liu, Jingwei Huang et al.

This paper proposes a 4D backbone for long-term point cloud video understanding. A typical way to capture spatial-temporal context is using 4Dconv or transformer without hierarchy. However, those methods are neither effective nor efficient enough due to camera motion, scene changes, sampling patterns, and the complexity of 4D data. To address those issues, we leverage the primitive plane as a mid-level representation to capture the long-term spatial-temporal context in 4D point cloud videos and propose a novel hierarchical backbone named Point Primitive Transformer(PPTr), which is mainly composed of intra-primitive point transformers and primitive transformers. Extensive experiments show that PPTr outperforms the previous state of the arts on different tasks.

Zhuoyang Zhang, Yuhao Dong, Yunze Liu et al.

Recent work on 4D point cloud sequences has attracted a lot of attention. However, obtaining exhaustively labeled 4D datasets is often very expensive and laborious, so it is especially important to investigate how to utilize raw unlabeled data. However, most existing self-supervised point cloud representation learning methods only consider geometry from a static snapshot omitting the fact that sequential observations of dynamic scenes could reveal more comprehensive geometric details. And the video representation learning frameworks mostly model motion as image space flows, let alone being 3D-geometric-aware. To overcome such issues, this paper proposes a new 4D self-supervised pre-training method called Complete-to-Partial 4D Distillation. Our key idea is to formulate 4D self-supervised representation learning as a teacher-student knowledge distillation framework and let the student learn useful 4D representations with the guidance of the teacher. Experiments show that this approach significantly outperforms previous pre-training approaches on a wide range of 4D point cloud sequence understanding tasks including indoor and outdoor scenarios.

Peiran Wu, Yunze Liu, Chi-Hao Wu et al.

Omnimodal large language models enable unified audio video understanding, but long joint token sequences make inference costly, and existing benchmarks do not fully isolate audio visual association in noisy user generated videos. We introduce UGC-AVQA, a public UGC benchmark with 1,000 videos and 4,816 QA pairs, where an audio removal test ensures that benchmark questions require both acoustic and visual evidence. To reduce inference cost, we propose OMAC, a training free plug in compression method that preserves salient visual memory and temporally grounded audio anchors. To further make compact models robust to compressed inputs, we introduce O-MARC, a compression distillation framework for learning with memory compressed multimodal contexts. On Qwen2.5-Omni-3B, O-MARC improves the average score across four benchmarks to 45.8, outperforming full token inference at 44.1 and OmniZip at 41.0. OMAC also keeps inference efficient, reducing latency by 34.6\% (1.53$\times$ speedup) and memory by 34.7\% compared with full token inference.

Jiazheng Li, Chi-Hao Wu, Yunze Liu et al.

Understanding ultra-long videos such as egocentric recordings, live streams, or surveillance footage spanning days to weeks, remains a challenge. For current multimodal LLMs: even with million-token context windows, frame budgets cover only tens of minutes of densely sampled video, and most evidence is discarded before inference begins. Memory-augmented and agentic approaches help with scale, but their retrieval remains fragmented across modalities and lacks long-range narrative summaries that span days or weeks. We propose \textbf{MAGIC-Video}, a training-free framework built around a multimodal memory graph with interleaved narrative chain: the graph unifies episodic, semantic, and visual content through six typed edges and supports cross-modal retrieval, while the chain distils long-horizon entity biographies and recurring activity events. At inference time, an agentic loop interleaves graph retrieval with narrative fact injection, covering both the modality and time dimensions of ultra-long video in a single retrieval pipeline. On EgoLifeQA, Ego-R1 and MM-Lifelong, MAGIC-Video consistently outperforms strong general-purpose, long-video, and agentic baselines, with gains of 10.1, 7.4, and 5.9 points over the prior best agentic system on each benchmark. Code is available at https://github.com/lijiazheng0917/MAGIC-video.

Yuhao Dong, Zhuoyang Zhang, Yunze Liu et al.

Understanding 4D point cloud sequences online is of significant practical value in various scenarios such as VR/AR, robotics, and autonomous driving. The key goal is to continuously analyze the geometry and dynamics of a 3D scene as unstructured and redundant point cloud sequences arrive. And the main challenge is to effectively model the long-term history while keeping computational costs manageable. To tackle these challenges, we introduce a generic online 4D perception paradigm called NSM4D. NSM4D serves as a plug-and-play strategy that can be adapted to existing 4D backbones, significantly enhancing their online perception capabilities for both indoor and outdoor scenarios. To efficiently capture the redundant 4D history, we propose a neural scene model that factorizes geometry and motion information by constructing geometry tokens separately storing geometry and motion features. Exploiting the history becomes as straightforward as querying the neural scene model. As the sequence progresses, the neural scene model dynamically deforms to align with new observations, effectively providing the historical context and updating itself with the new observations. By employing token representation, NSM4D also exhibits robustness to low-level sensor noise and maintains a compact size through a geometric sampling scheme. We integrate NSM4D with state-of-the-art 4D perception backbones, demonstrating significant improvements on various online perception benchmarks in indoor and outdoor settings. Notably, we achieve a 9.6% accuracy improvement for HOI4D online action segmentation and a 3.4% mIoU improvement for SemanticKITTI online semantic segmentation. Furthermore, we show that NSM4D inherently offers excellent scalability to longer sequences beyond the training set, which is crucial for real-world applications.

Yunze Liu, Peiran Wu, Cheng Liang et al.

Recent Mamba-based architectures for video understanding demonstrate promising computational efficiency and competitive performance, yet struggle with overfitting issues that hinder their scalability. To overcome this challenge, we introduce VideoMAP, a Hybrid Mamba-Transformer framework featuring a novel pre-training approach. VideoMAP uses a 4:1 Mamba-to-Transformer ratio, effectively balancing computational cost and model capacity. This architecture, combined with our proposed frame-wise masked autoregressive pre-training strategy, delivers significant performance gains when scaling to larger models. Additionally, VideoMAP exhibits impressive sample efficiency, significantly outperforming existing methods with less training data. Experiments show that VideoMAP outperforms existing models across various datasets, including Kinetics-400, Something-Something V2, Breakfast, and COIN. Furthermore, we demonstrate the potential of VideoMAP as a visual encoder for multimodal large language models, highlighting its ability to reduce memory usage and enable the processing of longer video sequences. The code is open-source at https://github.com/yunzeliu/MAP

Xinyi Zheng, Yunze Liu, Chi-Hao Wu et al.

We present SpatialMem, a memory-centric system that unifies 3D geometry, semantics, and language into a single, queryable representation. Starting from casually captured egocentric RGB video, SpatialMem reconstructs metrically scaled indoor environments, detects structural 3D anchors (walls, doors, windows) as the first-layer scaffold, and populates a hierarchical memory with open-vocabulary object nodes -- linking evidence patches, visual embeddings, and two-layer textual descriptions to 3D coordinates -- for compact storage and fast retrieval. This design enables interpretable reasoning over spatial relations (e.g., distance, direction, visibility) and supports downstream tasks such as language-guided navigation and object retrieval without specialized sensors. Experiments across three real-life indoor scenes demonstrate that SpatialMem maintains strong anchor-description-level navigation completion and hierarchical retrieval accuracy under increasing clutter and occlusion, offering an efficient and extensible framework for embodied spatial intelligence.

Xinyi Zheng, Steve Zhang, Weizhe Lin et al.

Current state-of-the-art 3D reconstruction models face limitations in building extra-large scale outdoor scenes, primarily due to the lack of sufficiently large-scale and detailed datasets. In this paper, we present a extra-large fine-grained dataset with 10 billion points composed of 41,006 drone-captured high-resolution aerial images, covering 20 diverse and culturally significant scenes from worldwide locations such as Cambridge Uni main buildings, the Pyramids, and the Forbidden City Palace. Compared to existing datasets, ours offers significantly larger scale and higher detail, uniquely suited for fine-grained 3D applications. Each scene contains an accurate spatial layout and comprehensive structural information, supporting detailed 3D reconstruction tasks. By reconstructing environments using these detailed images, our dataset supports multiple applications, including outputs in the widely adopted COLMAP format, establishing a novel benchmark for evaluating state-of-the-art large-scale Gaussian Splatting methods.The dataset's flexibility encourages innovations and supports model plug-ins, paving the way for future 3D breakthroughs. All datasets and code will be open-sourced for community use.

Yunze Liu, Zifan Wang, Peiran Wu et al.

Understanding dynamic 4D environments-3D space evolving over time-is critical for robotic and interactive systems. These applications demand systems that can process streaming point cloud video in real-time, often under resource constraints, while also benefiting from past and present observations when available. However, current 4D backbone networks rely heavily on spatiotemporal convolutions and Transformers, which are often computationally intensive and poorly suited to real-time applications. We propose PointNet4D, a lightweight 4D backbone optimized for both online and offline settings. At its core is a Hybrid Mamba-Transformer temporal fusion block, which integrates the efficient state-space modeling of Mamba and the bidirectional modeling power of Transformers. This enables PointNet4D to handle variable-length online sequences efficiently across different deployment scenarios. To enhance temporal understanding, we introduce 4DMAP, a frame-wise masked auto-regressive pretraining strategy that captures motion cues across frames. Our extensive evaluations across 9 tasks on 7 datasets, demonstrating consistent improvements across diverse domains. We further demonstrate PointNet4D's utility by building two robotic application systems: 4D Diffusion Policy and 4D Imitation Learning, achieving substantial gains on the RoboTwin and HandoverSim benchmarks.

Yunze Liu, Changxi Chen, Li Yi

We focus on the human-humanoid interaction task optionally with an object. We propose a new task named online full-body motion reaction synthesis, which generates humanoid reactions based on the human actor's motions. The previous work only focuses on human interaction without objects and generates body reactions without hand. Besides, they also do not consider the task as an online setting, which means the inability to observe information beyond the current moment in practical situations. To support this task, we construct two datasets named HHI and CoChair and propose a unified method. Specifically, we propose to construct a social affordance representation. We first select a social affordance carrier and use SE(3)-Equivariant Neural Networks to learn the local frame for the carrier, then we canonicalize the social affordance. Besides, we propose a social affordance forecasting scheme to enable the reactor to predict based on the imagined future. Experiments demonstrate that our approach can effectively generate high-quality reactions on HHI and CoChair. Furthermore, we also validate our method on existing human interaction datasets Interhuman and Chi3D.

Peiran Wu, Yunze Liu, Miao Liu et al.

Humans excel at spatial-temporal reasoning, effortlessly interpreting dynamic visual events from an egocentric viewpoint. However, whether multimodal large language models (MLLMs) can similarly understand the 4D world remains uncertain. This paper explores multimodal spatial-temporal reasoning from an egocentric perspective, aiming to equip MLLMs with human-like reasoning capabilities. To support this objective, we introduce \textbf{Ego-ST Bench}, a novel benchmark containing over 5,000 question-answer pairs across four categories, systematically evaluating spatial, temporal, and integrated spatial-temporal reasoning. Additionally, we propose \textbf{ST-R1} training paradigm, a video-based reasoning model that incorporates reverse thinking into its reinforcement learning process, significantly enhancing performance. We combine long-chain-of-thought (long-CoT) supervised fine-tuning with Group Relative Policy Optimization (GRPO) reinforcement learning, achieving notable improvements with limited high-quality data. Ego-ST Bench and ST-R1 provide valuable insights and resources for advancing video-based spatial-temporal reasoning research.

Yunze Liu, Changxi Chen, Chenjing Ding et al.

Humanoid Reaction Synthesis is pivotal for creating highly interactive and empathetic robots that can seamlessly integrate into human environments, enhancing the way we live, work, and communicate. However, it is difficult to learn the diverse interaction patterns of multiple humans and generate physically plausible reactions. The kinematics-based approaches face challenges, including issues like floating feet, sliding, penetration, and other problems that defy physical plausibility. The existing physics-based method often relies on kinematics-based methods to generate reference states, which struggle with the challenges posed by kinematic noise during action execution. Constrained by their reliance on diffusion models, these methods are unable to achieve real-time inference. In this work, we propose a Forward Dynamics Guided 4D Imitation method to generate physically plausible human-like reactions. The learned policy is capable of generating physically plausible and human-like reactions in real-time, significantly improving the speed(x33) and quality of reactions compared with the existing method. Our experiments on the InterHuman and Chi3D datasets, along with ablation studies, demonstrate the effectiveness of our approach.

Yunze Liu, Changxi Chen, Zifan Wang et al.

This paper introduces a novel approach named CrossVideo, which aims to enhance self-supervised cross-modal contrastive learning in the field of point cloud video understanding. Traditional supervised learning methods encounter limitations due to data scarcity and challenges in label acquisition. To address these issues, we propose a self-supervised learning method that leverages the cross-modal relationship between point cloud videos and image videos to acquire meaningful feature representations. Intra-modal and cross-modal contrastive learning techniques are employed to facilitate effective comprehension of point cloud video. We also propose a multi-level contrastive approach for both modalities. Through extensive experiments, we demonstrate that our method significantly surpasses previous state-of-the-art approaches, and we conduct comprehensive ablation studies to validate the effectiveness of our proposed designs.

Haowen Luo, Yunze Liu, Li Yi

The credibility and practicality of a reconstructed hand-object interaction sequence depend largely on its physical plausibility. However, due to high occlusions during hand-object interaction, physical plausibility remains a challenging criterion for purely vision-based tracking methods. To address this issue and enhance the results of existing hand trackers, this paper proposes a novel physically-aware hand motion de-noising method. Specifically, we introduce two learned loss terms that explicitly capture two crucial aspects of physical plausibility: grasp credibility and manipulation feasibility. These terms are used to train a physically-aware de-noising network. Qualitative and quantitative experiments demonstrate that our approach significantly improves both fine-grained physical plausibility and overall pose accuracy, surpassing current state-of-the-art de-noising methods.

Wenqi Zhou, Kai Cao, Hao Zheng et al.

Long-form egocentric video understanding provides rich contextual information and unique insights into long-term human behaviors, holding significant potential for applications in embodied intelligence, long-term activity analysis, and personalized assistive technologies. However, existing benchmark datasets primarily focus on single, short (\eg, minutes to tens of minutes) to moderately long videos, leaving a substantial gap in evaluating extensive, ultra-long egocentric video recordings. To address this, we introduce X-LeBench, a novel benchmark dataset meticulously designed to fill this gap by focusing on tasks requiring a comprehensive understanding of extremely long egocentric video recordings. Our X-LeBench develops a life-logging simulation pipeline that produces realistic, coherent daily plans aligned with real-world video data. This approach enables the flexible integration of synthetic daily plans with real-world footage from Ego4D-a massive-scale egocentric video dataset covers a wide range of daily life scenarios-resulting in 432 simulated video life logs spanning from 23 minutes to 16.4 hours. The evaluations of several baseline systems and multimodal large language models (MLLMs) reveal their poor performance across the board, highlighting the inherent challenges of long-form egocentric video understanding, such as temporal localization and reasoning, context aggregation, and memory retention, and underscoring the need for more advanced models.

Peiran Wu, Yunze Liu, Zhengdong Zhu et al.

Real-world user-generated videos, especially on platforms like TikTok, often feature rich and intertwined audio visual content. However, existing video captioning benchmarks and models remain predominantly visual centric, overlooking the crucial role of audio in conveying scene dynamics, speaker intent, and narrative context. This lack of omni datasets and lightweight, capable models hampers progress in fine grained, multimodal video understanding. To address these challenges, we introduce UGC-VideoCap, a new benchmark and model framework specifically designed for detailed omnimodal captioning of short form user-generated videos. Unlike prior datasets, UGC-VideoCap emphasizes balanced integration of audio and visual modalities, featuring 1000 TikTok videos annotated through a structured three stage human-in-the-loop pipeline covering audio only, visual only, and joint audio visual semantics. The benchmark also includes 4000 carefully crafted QA pairs probing both unimodal and cross modal understanding. Alongside the dataset, we propose UGC-VideoCaptioner(3B), a 3B parameter captioning model distilled from Gemini 2.5 Flash. Using a novel two-stage training strategy supervised fine tuning followed by Group Relative Policy Optimization (GRPO), our approach enables efficient adaptation from limited data while maintaining competitive performance. Together, our benchmark and model offer a high-quality foundation and a data-efficient solution for advancing omnimodal video captioning in unconstrained real-world UGC settings.

Peiran Wu, Zhuorui Yu, Yunze Liu et al.

The rapid progress of large language models (LLMs) has laid the foundation for multimodal models. However, visual language models (VLMs) still face heavy computational costs when extended from images to videos due to high frame rates and long durations. Token compression is a promising solution, yet most existing training-free methods cause information loss and performance degradation. To overcome this, we propose \textbf{Memory-Augmented Reinforcement Learning-based Token Compression (MARC)}, which integrates structured retrieval and RL-based distillation. MARC adopts a \textit{retrieve-then-compress} strategy using a \textbf{Visual Memory Retriever (VMR)} to select key clips and a \textbf{Compression Group Relative Policy Optimization (C-GRPO)} framework to distil reasoning ability from a teacher to a student model. Experiments on six video benchmarks show that MARC achieves near-baseline accuracy using only one frame's tokens -- reducing visual tokens by \textbf{95\%}, GPU memory by \textbf{72\%}, and latency by \textbf{23.9\%}. This demonstrates its potential for efficient, real-time video understanding in resource-constrained settings such as video QA, surveillance, and autonomous driving.

Yihong Ji, Yunze Liu, Yiyao Zhuo et al.

The perception and generation of Human-Object Interaction (HOI) are crucial for fields such as robotics, AR/VR, and human behavior understanding. However, current approaches model this task in an offline setting, where information at each time step can be drawn from the entire interaction sequence. In contrast, in real-world scenarios, the information available at each time step comes only from the current moment and historical data, i.e., an online setting. We find that offline methods perform poorly in an online context. Based on this observation, we propose two new tasks: Online HOI Generation and Perception. To address this task, we introduce the OnlineHOI framework, a network architecture based on the Mamba framework that employs a memory mechanism. By leveraging Mamba's powerful modeling capabilities for streaming data and the Memory mechanism's efficient integration of historical information, we achieve state-of-the-art results on the Core4D and OAKINK2 online generation tasks, as well as the online HOI4D perception task.

Yunze Liu, Qingnan Fan, Shanghang Zhang et al.

This paper proposes a method for representation learning of multimodal data using contrastive losses. A traditional approach is to contrast different modalities to learn the information shared between them. However, that approach could fail to learn the complementary synergies between modalities that might be useful for downstream tasks. Another approach is to concatenate all the modalities into a tuple and then contrast positive and negative tuple correspondences. However, that approach could consider only the stronger modalities while ignoring the weaker ones. To address these issues, we propose a novel contrastive learning objective, TupleInfoNCE. It contrasts tuples based not only on positive and negative correspondences but also by composing new negative tuples using modalities describing different scenes. Training with these additional negatives encourages the learning model to examine the correspondences among modalities in the same tuple, ensuring that weak modalities are not ignored. We provide a theoretical justification based on mutual information for why this approach works, and we propose a sample optimization algorithm to generate positive and negative samples to maximize training efficacy. We find that TupleInfoNCE significantly outperforms the previous state of the arts on three different downstream tasks.

Yunze Liu, Li Yi, Shanghang Zhang et al.

Self-supervised representation learning is a critical problem in computer vision, as it provides a way to pretrain feature extractors on large unlabeled datasets that can be used as an initialization for more efficient and effective training on downstream tasks. A promising approach is to use contrastive learning to learn a latent space where features are close for similar data samples and far apart for dissimilar ones. This approach has demonstrated tremendous success for pretraining both image and point cloud feature extractors, but it has been barely investigated for multi-modal RGB-D scans, especially with the goal of facilitating high-level scene understanding. To solve this problem, we propose contrasting "pairs of point-pixel pairs", where positives include pairs of RGB-D points in correspondence, and negatives include pairs where one of the two modalities has been disturbed and/or the two RGB-D points are not in correspondence. This provides extra flexibility in making hard negatives and helps networks to learn features from both modalities, not just the more discriminating one of the two. Experiments show that this proposed approach yields better performance on three large-scale RGB-D scene understanding benchmarks (ScanNet, SUN RGB-D, and 3RScan) than previous pretraining approaches.