Embodied AI Evidence Hub具身智能证据知识库

数据工程与质量 · Data Engineering and Quality

近一年 Ego-centric 数据手部检测与追踪的问题和难点

Challenges in Egocentric Hand Tracking

更新于

30 秒结论

先读这个判断

本调研关注 2025 年 7 月 29 日至 2026 年 7 月 29 日的第一视角数据，将“手部检测与追踪”定义为一条连续链路：从 2D 手框、关键点和左右手身份，到相机/世界坐标中的 3D 姿态、长时轨迹，再到手物接触与交互时刻。检索得到 172 篇去重候选，恢复 48 篇完整非 OCR 正文，对其中 15 篇进行深读和主张级校验。这是一次 scoping review，能回答近一年方法集中暴露的问题，但不代表所有会议和非 arXiv 工作的穷尽性计量。

精读论文
15
正式证据事件
29
知识单元
EA-DATA, EA-SENSOR, EA-HARDWARE, EA-4D
审计状态
当前结算 run 已通过仓库校验

v0 · 完整正文

研究综述

研究边界

本调研关注 2025 年 7 月 29 日至 2026 年 7 月 29 日的第一视角数据，将“手部检测与追踪”定义为一条连续链路：从 2D 手框、关键点和左右手身份，到相机/世界坐标中的 3D 姿态、长时轨迹，再到手物接触与交互时刻。检索得到 172 篇去重候选，恢复 48 篇完整非 OCR 正文，对其中 15 篇进行深读和主张级校验。这是一次 scoping review，能回答近一年方法集中暴露的问题，但不代表所有会议和非 arXiv 工作的穷尽性计量。

中心判断

近一年的关键变化，不是单帧手检测又多了一个更强骨干网络，而是研究对象从“当前帧里的手”转向“与相机运动解耦、在遮挡和出框期间仍可信的 4D 手轨迹”。实际瓶颈是三种结构性困难的乘积：观测性会突然归零，头部自运动会污染手的表观运动，而真实环境中精确 3D 真值又昂贵且难扩展。因此，只提高可见帧上的精度，不能自动换来更好的追踪；必须同时管理坐标系、观测质量、时序误差和真值偏差。

一、遮挡不是少量异常帧，而是长时追踪的常态

手在第一视角里频繁出框，手指又会被手掌、物体或另一只手遮挡。这不是“多加一点数据增强”就能消掉的尾部问题。DeltaDorsal 在 ARCTIC、H2O、EgoExo4D 和 AssemblyHands 上的统计显示，超过 20% 的帧至少有一根手指高度遮挡。手背皮肤形变可以在部分手指不可见时提供间接线索，但论文也明确限定了适用条件：手背本身要可见，并且需要足够高的分辨率。

更难的是长连续丢失。StableHand 把左/右手与腕部/手指的观测质量分开建模，说明单一“整手置信度”会掩盖局部可信、局部已失效的状态。然而，当两只手同时缺少可靠观测锚点时，生成先验只能补出“合理”动作，无法保证是真实轨迹。这条边界很重要：平滑不等于正确，时序先验不能创造已消失的观测信息。

二、头部自运动使“手动了多少”变成坐标系问题

在相机坐标中，头部抖动、身体行走和真实手动会叠加。EgoGrasp 因此将手物运动恢复到世界坐标，用场景几何、相机位姿和接触关系进行联合优化。这能极大改善稳定性，但其 200 步采样和测试时优化也表明，“稳定的离线恢复”还不等于可穿戴设备上的实时追踪。

换用事件相机也不会自动解决自运动。EgoEV-HandPose 明确指出，头部运动会在背景产生大量事件，与手的高频信号纠缠；单目事件流仍有深度歧义。双目几何能够改善深度，但同时引入双目标定、算力和功耗成本。即使是 RGB，EgoForce 也显示单目绝对 3D 依然受深度—尺度歧义、鱼眼/宽视场畸变和相机内参依赖影响。因而，相机模型不是可以在训练结束后再补的工程细节，而是任务定义的一部分。

三、时序上下文既是解药，也是误差放大器

直觉上，为追踪器增加记忆应当比逐帧预测更好。但 SFHand 给出了反例：素朴记忆会检索到背景 token，并把已写入队列的自回归误差在后续帧再次放大，表现甚至可以差于无记忆模型。这说明时序模型必须同时回答三个问题：记什么、何时忘、上游观测不可信时是否应拒绝更新。

时序目标越精细，标注误差越不能被忽略。TouchMoment 要在强头动、近距离遮挡与“几乎接触”之间找到精确的首次触碰帧，评测容差严格到 0–2 帧；但它的大规模自动训练标注与手工标签平均差 1.94 帧。也就是说，标注噪声与评测容差处在同一量级。对这类任务，光报一个平均精度远远不够，标注时间不确定性必须进入评测解释。

四、真实性、精确 3D 真值和采集成本无法同时最优

第一视角数据有一个基本张力：越是真实、自然、可移动的场景，越难得到毫米级 3D 手真值。Ego–Exo 移动采集架 使用 8 个 exo 鱼眼相机、2 个 ego 相机、OptiTrack、硬件同步与标定，才将可验证的 3D 手标注带到移动场景。这种系统有力，但 8 公斤级采集装置本身也会改变参与者的行为和场景覆盖。

合成数据能扩展姿态和标注，但不能被当成真实数据的无条件替代。HOI-Synth 的跨数据集实验表明，合成数据在少量真实标注和域适应下有价值，但纯合成数据不能完全替代真实 HOI 数据；而且该证据是单帧分割/检测，并没有证明轨迹连续性。EvHand-FPV 的现实边界更直接：真实事件数据只有 2D 标注，因此真实 3D 性能无法直接验证，论文的 3D 评测只能留在合成域。

手物接触数据更难。EPIC-Contact / HOPformer 用中心帧人工接触标注向整段 clip 传播，扩大了野外 3D HOI 监督，但传播会被左右手辨识错误和相机跳变污染。这提示数据集不应只发布一个硬标签，还要保留生成路径、每帧置信度和可能的传播失效。

五、多模态不是免费的精度奖励

事件相机对快速运动、低光和低延迟有价值，但它对慢变信号不敏感，将事件流重新帧化又会损失异步表示的优势。EventEgoHands 用 RGB 视频经 v2e 合成事件，并主要依赖伪标签扩大手框数据；它对手检测有直接价值，但并不覆盖真实事件噪声、3D、左右手身份或长时追踪。

视觉–IMU 融合可以在遮挡下保留手指的高频运动，同时由视觉给腕部绝对位置提供锚点。AVI-HT 支持这种互补性，但也报告了对 IMU 噪声和时间对齐的敏感性。同时，手套会改变手的外观；AirGlove 表明，在裸手上预训练的追踪器遇到不同传感手套时性能会大幅下降。传感器增加的不只是信息，也有同步、标定、功耗、外观域差和跨设备泛化债务。

六、当前评测容易高估“真实可用的追踪”

一些方法已在可见帧上做到实时。例如 Hand-4DGS 报告约 60 FPS 的前馈式 4D 手重建。但它的定量评测排除了超过一半手框出界，或上游姿态器无法正确检测双手的帧。这不否定其在可评估区间的进步，但说明“成功帧内的 MPJPE”不能代表端到端可用性。

对追踪系统，评测至少要分开四层：可检测率，可见期间的 2D/3D 精度，遮挡或出框后的身份续接/重获取，以及在世界坐标中的长时漂移。如果一个数据集只标手框，它可以验证检测，但不能据此声称解决追踪；如果只在合成数据上有 3D 真值，真实 3D 性能就必须保留为未验证。

面向系统设计的可操作框架

层级必须保留的数据主指标压力测试不应被平均掉的失败
检测全帧手框/掩码、出框状态recall、mAP、连续丢失长度模糊、强自运动、低光、双手交叉未检测帧被先排除
姿态2D/3D 关键点、相机内参、标尺MPJPE/PA-MPJPE、腕部绝对误差自遮挡、手物遮挡、鱼眼、跨设备相对姿态很好但绝对深度错
追踪身份、时间戳、观测质量、遮挡区间ID 续接、重获取延迟、漂移、加速误差长时出框、双手交换、错误记忆回灌先验补出平滑但错误的轨迹
接触/HOI手物接触、物体位姿、标注不确定性触碰时刻误差、接触偏差、交互成功近接触、手边别错误、相机跳变轮廓吻合却接触关系错

研究空白与下一步

论文直接支持的空白有三个：真实事件数据的 3D 真值缺失；野外手物 3D 接触监督仍难以便宜扩展；新事件手数据集的光照、肤色和活动覆盖仍不足。

基于多篇证据，本文进一步推断，最值得投资的研究方向是“失败可观测的追踪”：模型应当显式输出左/右手、腕/指、位置/姿态的分解不确定性；在无观测锚点时区分“预测”和“观测”；评测从完整时间线起算，包括丢失和重获取。这个推断可被反证：如果一个仅优化可见帧单帧精度的模型，在未剪裁的真实长视频上同时显著降低身份中断、重获取延迟和世界轨迹漂移，那么显式失败建模的必要性就会减弱。

结论

Egocentric 手部追踪的实质已不再是一个独立检测器的问题，而是观测性、坐标系、时序不确定性、真值生产和端侧成本的联合问题。近一年的方法已经对遮挡恢复、世界坐标稳定、事件/惯性融合和野外接触监督提出了有效工具；但它们共同揭示的更重要结论是：没有任何一种新传感器或时序先验能无条件填补消失的观测。下一阶段应优先让数据和评测忠实保留失败，再让模型学会在失败中追踪。

References

  1. EvHand-FPV: Efficient Event-Based 3D Hand Tracking from First-Person View
  2. Towards Lightweight and Mobile Ground Truth Systems for Ego-Exo 3D Hand Tracking
  3. SFHand: Language-Guided Streaming 3D Hand Forecasting
  4. EgoGrasp: World-Space Hand-Object Interaction Recovery
  5. DeltaDorsal: Hand Pose from Dorsal Skin Deformation
  6. AirGlove: Appearance-Invariant Hand Tracking for Sensing Gloves
  7. Leveraging Synthetic Data for Enhancing Egocentric Hand-Object Interaction Detection
  8. Detecting Precise Hand Touch Moments in Egocentric Video
  9. EgoEV-HandPose: Egocentric 3D Hand Pose Estimation and Gesture Recognition with Stereo Event Cameras
  10. EgoForce: Forearm-Guided Camera-Space 3D Hand Pose
  11. StableHand: Quality-Aware Flow Matching for World-Space Dual-Hand Motion Estimation
  12. AVI-HT: Adaptive Vision-IMU Fusion for 3D Hand Tracking
  13. A Multimodal RGB and Events Dataset for Hand Detection in First-Person View
  14. Hand-4DGS: Feed-Forward 3D Gaussian Splatting for 4D Hand Reconstruction
  15. Towards in-the-wild Egocentric 3D Hand-Object Pose Estimation

Accepted evidence

去重论文引用

  1. A Multimodal RGB and Events Dataset for Hand Detection in First-Person ViewBharghav Kota, Yulia Sandamirskaya · 2026
  2. AirGlove: Exploring Egocentric 3D Hand Tracking and Appearance Generalization for Sensing GlovesWenhui Cui, Ziyi Kou, Chuan Qin, Ergys Ristani, Li Guan · 2026
  3. AVI-HT: Adaptive Vision-IMU Fusion for 3D Hand TrackingZiyi Kou, Ankit Kumar, Mia Huang, Taylor Niehues, Vatsal Mehta, Ergys Ristani, Li Guan · 2026
  4. DeltaDorsal: Enhancing Hand Pose Estimation with Dorsal Features in Egocentric ViewsWilliam Huang, Siyou Pei, Leyi Zou, Eric J. Gonzalez, Ishan Chatterjee, Yang Zhang · 2026
  5. Detecting Precise Hand Touch Moments in Egocentric VideoHuy Anh Nguyen, Feras Dayoub, Minh Hoai · 2026
  6. Ego-Exo 3D Hand Tracking in the Wild with a Mobile Multi-Camera RigPatrick Rim, Kun He, Kevin Harris, Braden Copple, Shangchen Han, Sizhe An, Ivan Shugurov, Tomas Hodan et al. · 2025
  7. EgoEV-HandPose: Egocentric 3D Hand Pose Estimation and Gesture Recognition with Stereo Event CamerasLuming Wang, Hao Shi, Jiajun Zhai, Kailun Yang, Kaiwei Wang · 2026
  8. EgoForce: Forearm-Guided Camera-Space 3D Hand Pose from a Monocular Egocentric CameraChristen Millerdurai, Shaoxiang Wang, Yaxu Xie, Vladislav Golyanik, Didier Stricker, Alain Pagani · 2026
  9. EgoGrasp: World-Space Hand-Object Interaction Estimation from Egocentric VideosHongming Fu, Wenjia Wang, Xiaozhen Qiao, Rolandos Alexandros Potamias, Taku Komura, Shuo Yang, Zheng Liu, Bo Zhao · 2026
  10. EvHand-FPV: Efficient Event-Based 3D Hand Tracking from First-Person ViewZhen Xu, Guorui Lu, Chang Gao, Qinyu Chen · 2025
  11. Hand-4DGS: Feed-Forward 3D Gaussian Splatting for 4D Hand Reconstruction from Egocentric VideosJeongmin Bae, Seoha Kim, Marc Pollefeys, Mahdi Rad, Youngjung Uh, Taein Kwon · 2026
  12. Leveraging Synthetic Data for Enhancing Egocentric Hand-Object Interaction DetectionRosario Leonardi, Antonino Furnari, Francesco Ragusa, Giovanni Maria Farinella · 2026
  13. SFHand: Learning Embodied Manipulation by Streaming Egocentric 3D Hand ForecastingRuicong Liu, Yifei Huang, Liangyang Ouyang, Caixin Kang, Yoichi Sato · 2025
  14. StableHand: Quality-Aware Flow Matching for World-Space Dual-Hand Motion Estimation from Egocentric VideoHuajian Zeng, Chaohua Yao, Yuantai Zhang, Jiaqi Yang, Rolandos Alexandros Potamias, Xingxing Zuo · 2026
  15. Towards in-the-wild Egocentric 3D Hand-Object Pose EstimationSiddhant Bansal, Zhifan Zhu, Shashank Tripathi, Jiahe Zhao, Michael J. Black, Dima Damen · 2026
完整证据附录（29 条正式证据事件）
  • Time range: 2025-07-29 至 2026-07-29
  • Events: 29
  • 每个事件一节,标题即锚点;trace-map 中的 event 链接跳转到这里。

EA-EGOHAND-2026-0007

  • Claim: 第一视角的相机坐标会把头部抖动与物体运动混合，稳定 HOI 追踪需要世界坐标锚定。
  • Stance: support | Confidence: direct
  • Paper: 2601.01050 EgoGrasp: World-Space Hand-Object Interaction Estimation from Egocentric Videos
  • Locator: 1 Introduction
  • Evidence: 原文直接描述相机自运动与物体运动的坐标耦合。
  • Quote: “In egocentric scenarios, high-frequency camera jitter is entangled with object movement in local views.”
  • Authors: hongming-fu; wenjia-wang; xiaozhen-qiao; et al.

EA-EGOHAND-2026-0009

  • Claim: Egocentric 手指自遮挡是高频现象：跨四个数据集，超过 20% 帧至少有一根手指高度遮挡。
  • Stance: support | Confidence: direct
  • Paper: 2601.15516 DeltaDorsal: Enhancing Hand Pose Estimation with Dorsal Features in Egocentric Views
  • Locator: 3.2. Occlusion Prevalence in Egocentric Data
  • Evidence: 原文在四数据集分析中直接报告遮挡比例。
  • Quote: “As shown in Figure 3 a, more than 20% of frames across all evaluated datasets exhibit at least one occluded finger, with over 5% containing two or more occluded fingers.”
  • Authors: william-huang; siyou-pei; leyi-zou; et al.

EA-EGOHAND-2026-0018

  • Claim: 单目头戴相机中，绝对 3D 手追踪同时受深度–尺度歧义、自遮挡与宽 FOV/鱼眼变形限制。
  • Stance: support | Confidence: direct
  • Paper: 2605.12498 EgoForce: Forearm-Guided Camera-Space 3D Hand Pose from a Monocular Egocentric Camera
  • Locator: 1. Introduction
  • Evidence: 原文在问题定义中直接并列了绝对 3D 手追踪的三个主要视觉/几何难点。
  • Quote: “However, achieving this from a single egocentric camera is challenging due to depth–scale ambiguity, frequent self-occlusions, and the strong distortions introduced by wide-FOV and fisheye optics (Millerdurai et al. , 2024a , 2025 ) .”
  • Authors: christen-millerdurai; shaoxiang-wang; yaxu-xie; et al.

EA-EGOHAND-2026-0020

  • Claim: Ego 双手追踪的观测质量具有左/右手与腕部/手指两个轴的异质性，不能用单一置信度概括。
  • Stance: support | Confidence: direct
  • Paper: 2605.18553 StableHand: Quality-Aware Flow Matching for World-Space Dual-Hand Motion Estimation from Egocentric Video
  • Locator: 1 Introduction
  • Evidence: 原文直接给出单一质量分数会混淆的两类异质性。
  • Quote: “A single scalar quality per hand would average wrist global drift with finger articulation error and ignore asymmetric bimanual occlusion, informing our four-channel signal indexed by hand and component.”
  • Authors: huajian-zeng; chaohua-yao; yuantai-zhang; et al.

EA-EGOHAND-2026-0026

  • Claim: Egocentric 4D 手重建同时受移动相机、严重自遮挡、有限视角、快速手动与双手交互影响。
  • Stance: support | Confidence: direct
  • Paper: 2606.19156 Hand-4DGS: Feed-Forward 3D Gaussian Splatting for 4D Hand Reconstruction from Egocentric Videos
  • Locator: 1 Introduction
  • Evidence: 原文在问题定义中直接列出了移动相机、遮挡、视角、快速运动和双手交互等耦合困难。
  • Quote: “Body reconstruction methods typically assume a static camera observing a single person, whereas egocentric scenarios involve a moving camera, severe self-occlusions, limited viewpoints, rapid hand motion, and interactions between two hands.”
  • Authors: jeongmin-bae; seoha-kim; marc-pollefeys; et al.

EA-EGOHAND-2026-0004

  • Claim: 在真实环境中获得精确 3D 手真值仍依赖多相机、同步、标定和重型移动采集硬件。
  • Stance: conditional | Confidence: direct
  • Paper: 2510.02601 Ego-Exo 3D Hand Tracking in the Wild with a Mobile Multi-Camera Rig
  • Locator: 3.1 Mobile Capture Rig
  • Evidence: 原文的硬件章节表明高精度真值来自多视图、MoCap、硬同步和标定的组合。
  • Quote: “All ten monochrome cameras and the five MoCap cameras are hardware-synchronized and precisely calibrated into a shared three-dimensional reference frame.”
  • Authors: patrick-rim; kun-he; kevin-harris; et al.

EA-EGOHAND-2026-0008

  • Claim: 开放词表世界坐标 HOI 恢复的精度目前以重型离线管线为代价，不等于可实时部署。
  • Stance: conditional | Confidence: direct
  • Paper: 2601.01050 EgoGrasp: World-Space Hand-Object Interaction Estimation from Egocentric Videos
  • Locator: 4.1 Implementation Details & Metrics
  • Evidence: 推理设置明确使用 200 步扩散采样，且前文还有多步测试时优化。
  • Quote: “For inference, we applied DDIM [ song2020denoising ] sampling with 200 steps, and downsampled sequences by 3 at preprocessing stage.”
  • Authors: hongming-fu; wenjia-wang; xiaozhen-qiao; et al.

EA-EGOHAND-2026-0010

  • Claim: 手背皮肤形变只是条件性遮挡补充信号；手背不可见、低分辨率或快速运动会削弱其价值。
  • Stance: conditional | Confidence: direct
  • Paper: 2601.15516 DeltaDorsal: Enhancing Hand Pose Estimation with Dorsal Features in Egocentric Views
  • Locator: 8. Discussion
  • Evidence: 作者在泛化讨论中明确限定手背特征的可见性条件。
  • Quote: “Secondly, we acknowledge that our approach leveraging purely dorsal features for pose estimation is only applicable for a portion of in-the-wild scenarios where the dorsum of the hand is visible.”
  • Authors: william-huang; siyou-pei; leyi-zou; et al.

EA-EGOHAND-2026-0012

  • Claim: 当前纯合成数据不能替代真实 egocentric HOI 数据；它主要是少标签和域适应下的补充资源。
  • Stance: conditional | Confidence: direct
  • Paper: 2603.29733 Leveraging Synthetic Data for Enhancing Egocentric Hand-Object Interaction Detection
  • Locator: 5 Discussion
  • Evidence: 讨论章直接给出纯合成方案的边界。
  • Quote: “Currently, synthetic data alone cannot fully replace real-world data for egocentric HOI detection.”
  • Authors: rosario-leonardi; antonino-furnari; francesco-ragusa; et al.

EA-EGOHAND-2026-0015

  • Claim: TouchMoment 的自动训练标注与手工标签平均相差 1.94 帧，这与严格评测容差处于同一量级。
  • Stance: conditional | Confidence: direct
  • Paper: 2604.12343 Detecting Precise Hand Touch Moments in Egocentric Video
  • Locator: 3.2 Touch Annotation
  • Evidence: 原文给出自动工具相对手工标注的帧级差异。
  • Quote: “On a held-out subset with full manual annotation, this automatic tool differs from manual labels by an average of 1.94 frames, demonstrating sufficient accuracy for large-scale training data.”
  • Authors: huy-anh-nguyen; feras-dayoub; minh-hoai

EA-EGOHAND-2026-0017

  • Claim: 双目事件的几何增益以更高硬件、标定和算力成本为代价，超低功耗部署仍需压缩。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.12297 EgoEV-HandPose: Egocentric 3D Hand Pose Estimation and Gesture Recognition with Stereo Event Cameras
  • Locator: V-I Limitations
  • Evidence: 作者在限制章明确承认当前架构对超低功耗部署仍有额外工程需求。
  • Quote: “To facilitate deployment on ultra-low-power wearable devices with stringent power budgets, such overhead can be further mitigated through model compression or quantization techniques.”
  • Authors: luming-wang; hao-shi; jiajun-zhai; et al.

EA-EGOHAND-2026-0022

  • Claim: 视觉与 6-DoF IMU 的互补只有在准确同步下成立：视觉锚定全局位置，IMU 补足遮挡下的高频指部运动。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.21714 AVI-HT: Adaptive Vision-IMU Fusion for 3D Hand Tracking
  • Locator: 1 Introduction
  • Evidence: 该句在上下文中明确指向视觉全局锚点与 IMU 局部动力学的互补。
  • Quote: “Therefore, fusing the two recovers what each modality alone cannot.”
  • Authors: ziyi-kou; ankit-kumar; mia-huang; et al.

EA-EGOHAND-2026-0001

  • Claim: 事件手追踪存在显著视角域差：第三视角模型直接用于第一视角时性能会严重下降。
  • Stance: limit | Confidence: direct
  • Paper: 2509.13883 EvHand-FPV: Efficient Event-Based 3D Hand Tracking from First-Person View
  • Locator: IV-B Comparison with Prior Work
  • Evidence: 原文在同一表中报告了 EventHands 直接跨视角迁移的下降。
  • Quote: “However, when applied to our first-person perspective (egocentric) task, its performance drops dramatically to only 0.12 2D-AUCp on real data and 0.17 3D-AUC on synthetic data, demonstrating the significant domain gap between viewpoints and the necessity of EvHand-FPV.”
  • Authors: zhen-xu; guorui-lu; chang-gao; et al.

EA-EGOHAND-2026-0003

  • Claim: 受控数据上训练的 3D 手追踪器不能保证野外泛化；在 EgoExo-Hands 上的 MKPE 从域内约 9–11 mm 上升到 16.28 mm。
  • Stance: limit | Confidence: direct
  • Paper: 2510.02601 Ego-Exo 3D Hand Tracking in the Wild with a Mobile Multi-Camera Rig
  • Locator: 3.3 Quantitative Evaluation
  • Evidence: 该结论由同节 Table 2 的域内和跨域 MKPE 直接支持。
  • Quote: “Models trained on existing datasets, UmeTrack [ 9 ] and HOT3D [ 2 ] , generalize significantly worse to EgoExo-Hands, highlighting the increased difficulty of our in-the-wild data.”
  • Authors: patrick-rim; kun-he; kevin-harris; et al.

EA-EGOHAND-2026-0005

  • Claim: 流式手追踪/预测中，素朴时序记忆可能比无记忆更差，因为背景 token 会检索并放大历史队列中的自回归误差。
  • Stance: limit | Confidence: direct
  • Paper: 2511.18127 SFHand: Learning Embodied Manipulation by Streaming Egocentric 3D Hand Forecasting
  • Locator: 5.3.2 Ablation on ROI-enhanced memory
  • Evidence: 该句由同节 Table 4 的 ADE/FDE 消融数据直接支持。
  • Quote: “The vanilla memory without ROI enhancement performs even worse than the no-memory baseline.”
  • Authors: ruicong-liu; yifei-huang; liangyang-ouyang; et al.

EA-EGOHAND-2026-0006

  • Claim: 自回归追踪/预测必须显式管理自身误差累积；被存入的错误可以被后续时刻再次检索并放大。
  • Stance: limit | Confidence: direct
  • Paper: 2511.18127 SFHand: Learning Embodied Manipulation by Streaming Egocentric 3D Hand Forecasting
  • Locator: 5.3.2 Ablation on ROI-enhanced memory
  • Evidence: 作者在消融解释中直接给出了记忆导致误差放大的机制。
  • Quote: “This can amplify error, as autoregressive prediction errors stored in the memory queue ( ) are retrieved and propagated by non-salient parts of the current query ( ).”
  • Authors: ruicong-liu; yifei-huang; liangyang-ouyang; et al.

EA-EGOHAND-2026-0011

  • Claim: 裸手预训练的视觉手追踪器在传感手套上存在大幅外观域差，每种新手套都可能需要新的适配数据。
  • Stance: limit | Confidence: direct
  • Paper: 2602.05159 AirGlove: Exploring Egocentric 3D Hand Tracking and Appearance Generalization for Sensing Gloves
  • Locator: 5 Evaluation
  • Evidence: 原文在四类手套评测表后直接总结了裸手到手套的性能下降。
  • Quote: “We observe that both MEgATrack and UmeTrack show substantially degraded performance on sensing gloves compared to bare-hands.”
  • Authors: wenhui-cui; ziyi-kou; chuan-qin; et al.

EA-EGOHAND-2026-0013

  • Claim: HOI-Synth 的证据是单帧检测证据，不包含轨迹连续性或时序追踪能力。
  • Stance: limit | Confidence: direct
  • Paper: 2603.29733 Leveraging Synthetic Data for Enhancing Egocentric Hand-Object Interaction Detection
  • Locator: 5.1 Limitations and future work
  • Evidence: 限制章明确声明不使用时序信息。
  • Quote: “Our analysis is restricted to frame-level HOS detection and does not exploit temporal information.”
  • Authors: rosario-leonardi; antonino-furnari; francesco-ragusa; et al.

EA-EGOHAND-2026-0014

  • Claim: 精确触碰时刻检测不是普通手检测；它还需区分强自运动、近距遮挡和视觉上几乎相同的近接触帧。
  • Stance: limit | Confidence: direct
  • Paper: 2604.12343 Detecting Precise Hand Touch Moments in Egocentric Video
  • Locator: 1 Introduction
  • Evidence: 原文直接列出了接触时刻定位中的第一视角特有干扰。
  • Quote: “Rapid head motion introduces strong egomotion, while approaching hands often create severe occlusions and perspective distortions at close range.”
  • Authors: huy-anh-nguyen; feras-dayoub; minh-hoai

EA-EGOHAND-2026-0016

  • Claim: 事件相机不会自动消除 egocentric 干扰：头部运动生成的背景事件会与手运动信号耦合。
  • Stance: limit | Confidence: direct
  • Paper: 2605.12297 EgoEV-HandPose: Egocentric 3D Hand Pose Estimation and Gesture Recognition with Stereo Event Cameras
  • Locator: I Introduction
  • Evidence: 原文直接描述了头部自运动对事件流的混入机制。
  • Quote: “However, processing such asynchronous event video streams in egocentric settings is fundamentally challenged by ego-motion noise: continuous head movements induce background events that couple with hand-generated signals and severely complicate feature decoupling [ 22 , 34 ] .”
  • Authors: luming-wang; hao-shi; jiajun-zhai; et al.

EA-EGOHAND-2026-0019

  • Claim: 跨镜头单网络仍依赖已标定 3D 训练数据和相机内参，不等于无标定野外泛化。
  • Stance: limit | Confidence: direct
  • Paper: 2605.12498 EgoForce: Forearm-Guided Camera-Space 3D Hand Pose from a Monocular Egocentric Camera
  • Locator: 5. Limitations
  • Evidence: 作者在限制章明确把已标定 3D 训练依赖与野外泛化不足联系起来。
  • Quote: “It relies on calibrated 3D datasets for training, preventing the use of large 2D hand datasets common in root-relative methods (Pavlakos et al. , 2024 ; Potamias et al. , 2024 ) and limiting generalization to in-the-wild imagery.”
  • Authors: christen-millerdurai; shaoxiang-wang; yaxu-xie; et al.

EA-EGOHAND-2026-0021

  • Claim: 当上游视觉完全没有可靠观测锚点时，生成恢复只能产生合理先验，可能形成错误轨迹。
  • Stance: limit | Confidence: direct
  • Paper: 2605.18553 StableHand: Quality-Aware Flow Matching for World-Space Dual-Hand Motion Estimation from Egocentric Video
  • Locator: Appendix H Additional Qualitative Results
  • Evidence: 作者在失败案例中明确区分了“合理”与“正确”。
  • Quote: “In both cases the generative process synthesizes a plausible but incorrect trajectory from the prior alone, illustrating the limit of our method when the upstream visual stream provides no observation to anchor a channel.”
  • Authors: huajian-zeng; chaohua-yao; yuantai-zhang; et al.

EA-EGOHAND-2026-0023

  • Claim: 手套型多模态追踪会引入新的外观域差，且对不同手套布局和 IMU 规格的泛化尚未验证。
  • Stance: limit | Confidence: direct
  • Paper: 2605.21714 AVI-HT: Adaptive Vision-IMU Fusion for 3D Hand Tracking
  • Locator: 6 Conclusion
  • Evidence: 原文在结论的限制段明确将跨手套泛化列为未验证问题。
  • Quote: “In addition, since our experiments use a specific type of data sensing glove, the generalization to other gloves with different sensor layouts, form factors, or IMU specifications remains to be validated.”
  • Authors: ziyi-kou; ankit-kumar; mia-huang; et al.

EA-EGOHAND-2026-0024

  • Claim: 事件手检测的精度仍受慢变信号不可见与将事件流重新帧化的处理低效限制。
  • Stance: limit | Confidence: direct
  • Paper: 2606.10790 A Multimodal RGB and Events Dataset for Hand Detection in First-Person View
  • Locator: I Introduction
  • Evidence: 原文直接列出了事件传感器与处理表示两类瓶颈。
  • Quote: “But event-based approaches currently face limitations in accuracy due to two main factors: the sensors’ inability to detect slowly changing signals and the inefficiency of existing processing techniques that transform event streams into frame- based formats for analysis using convolutional neural networks [ 31 , 22 , 1 ] .”
  • Authors: bharghav-kota; yulia-sandamirskaya

EA-EGOHAND-2026-0027

  • Claim: Hand-4DGS 的定量结果排除了手大部分出框或上游无法正确检测双手的帧，因而不能外推到最难丢检场景。
  • Stance: limit | Confidence: direct
  • Paper: 2606.19156 Hand-4DGS: Feed-Forward 3D Gaussian Splatting for 4D Hand Reconstruction from Egocentric Videos
  • Locator: C.2 Training Setup and Model Architecture
  • Evidence: 数据处理细节明确列出了两类被排除的高难样本。
  • Quote: “We perform training and evaluation only when both hands are sufficiently visible in the image. Specifically, we exclude instances where more than half of a bounding box area is outside the image or when HaMeR fails to detect both hands correctly.”
  • Authors: jeongmin-bae; seoha-kim; marc-pollefeys; et al.

EA-EGOHAND-2026-0029

  • Claim: 将中心帧的接触标注传播到整个 clip 会受手边别错误和相机跳变污染，需保留每帧置信度。
  • Stance: limit | Confidence: direct
  • Paper: 2606.30598 Towards in-the-wild Egocentric 3D Hand-Object Pose Estimation
  • Locator: 10 Limitations and Future Directions
  • Evidence: 作者在限制中明确列出了标签传播的两类时序噪声。
  • Quote: “Additionally, we propagate manually-verified ground-truth from a single frame to a clip. Propagation can be noisy due to errors in hand pose estimates from WiLoR [ Potamias_2025_CVPR_wilor ] (hand-side errors or camera placement jumps).”
  • Authors: siddhant-bansal; zhifan-zhu; shashank-tripathi; et al.

EA-EGOHAND-2026-0002

  • Claim: 真实事件数据缺少 3D 真值，导致该方法的真实 3D 指标无法被直接验证。
  • Stance: gap | Confidence: direct
  • Paper: 2509.13883 EvHand-FPV: Efficient Event-Based 3D Hand Tracking from First-Person View
  • Locator: IV-A2 3D Metric
  • Evidence: 原文明确将 3D 评测限定在合成数据。
  • Quote: “However, due to the inherent difficulties in obtaining 3D annotations for real-world data, our evaluation of this metric is limited to synthetic data.”
  • Authors: zhen-xu; guorui-lu; chang-gao; et al.

EA-EGOHAND-2026-0025

  • Claim: 现有 EventEgoHands 仍缺少光照、肤色和活动多样性，真实数据覆盖是未解问题。
  • Stance: gap | Confidence: direct
  • Paper: 2606.10790 A Multimodal RGB and Events Dataset for Hand Detection in First-Person View
  • Locator: V conclusion and further work
  • Evidence: 未来工作明确列出了数据人群与场景多样性边界。
  • Quote: “Future work includes creating an event-based Hands dataset in first-person view with more variation in lighting conditions, skin tones and activity being performed.”
  • Authors: bharghav-kota; yulia-sandamirskaya

EA-EGOHAND-2026-0028

  • Claim: 野外 egocentric 手物 3D 估计的主要数据瓶颈是重遮挡与接触歧义下缺少便宜、可扩展的 3D 监督。
  • Stance: gap | Confidence: direct
  • Paper: 2606.30598 Towards in-the-wild Egocentric 3D Hand-Object Pose Estimation
  • Locator: 1 Introduction
  • Evidence: 该句所在段落将野外遮挡/接触歧义与 MoCap 成本、环境空洞联系起来。
  • Quote: “A central bottleneck is the availability of supervision.”
  • Authors: siddhant-bansal; zhifan-zhu; shashank-tripathi; et al.

References

Interactive Wiki

进入交互阅读

静态页采用默认版本；v0、v0、v0可在同一交互 Wiki 中阅读，不建立相互竞争的索引页。

用 AI 继续研究 v0 v0 v0

Method and limits

研究方法、证据边界与引用说明

本页由当前结算的证据层自动派生，不改写已结算 run。论文只有在完整非 OCR 全文、结构化论文笔记和 claim-support audit 通过后，才进入接受证据。

页面结论是对所列论文与研究窗口的综合，不等同于无限外推。引用本专题时，请同时保留本页 canonical 地址、访问日期，并在需要时回溯原论文。