Embodied AI Evidence Hub具身智能证据知识库

综合研究 · Embodied AI Research

Ego-Exo 后继研究：第三人称→第一人称视觉表征迁移的演进

Ego–Exo Representation Transfer

更新于

30 秒结论

先读这个判断

研究备忘录 主题：第三人称→第一人称视觉表征迁移的演进 证据基础：11 篇全文精读论文，20 条经验证证据事件 综述模式：scoping 本备忘录基于经过完整全文提取与 claim-support audit 的论文级证据，所有主张可追溯到具体论文。事件级溯源见 evidence-appendix.md 与 trace-map.json。

精读论文
0
正式证据事件
20
知识单元
EA-XEMBODIMENT, EA-MODEL, EA-DATA
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

研究备忘录 | 主题：第三人称→第一人称视觉表征迁移的演进 | 证据基础：11 篇全文精读论文，20 条经验证证据事件 | 综述模式：scoping 本备忘录基于经过完整全文提取与 claim-support audit 的论文级证据，所有主张可追溯到具体论文。事件级溯源见 evidence-appendix.md 与 trace-map.json。

1. 研究边界与问题界定

本文回答一个具体问题：Ego-Exo（Li et al., CVPR 2021）提出的"从第三人称视频迁移视觉表征到第一人称"思路，其后继工作如何演进，形成了哪些子方向与数据/方法范式？

证据池由引文分析确定：以 Ego-Exo 为根节点，经 Semantic Scholar 引文图向下游扩展并人工复核，选出 11 篇真正属于 ego-exo / third-person / exo 领域的后继论文，全部完成非 OCR 全文提取与六遍式精读，投影出 20 条证据事件。需要说明：这是一个引文图派生的候选集，不是关键词全量检索，因此不声称覆盖该领域全部工作，也不把"被引影响力"等同于"方法正确性"。

2. 中心判断

Ego-Exo 的真正贡献不是某一个模型，而是把一个直觉——"第三人称视频里藏着能被蒸馏到第一人称模型的 egocentric 信号"——变成了一个可复用的迁移范式。十年间，这个范式沿三条路线展开：(1) 数据/基准化（把视角迁移固化为大规模同步或异步 ego-exo 数据集）；(2) 任务化（把"迁移"落到 affordance 定位、对象对应、时序动作分割、视频描述等具体下游任务）；(3) 前置放松（从"无需配对/标注 ego 视频"进一步到"无需同步""无需配对""用语言/gaze 桥接"）。

贯穿这三条路线的核心张力是：难点从来不在"要不要迁移"，而在"用什么信号桥接视角差距"。Ego-Exo 用自动伪标签，后续工作依次验证了时间同步、gaze、语言对齐、无标注同步视频对作为桥接信号的有效性；而一旦桥接信号缺失或过弱，迁移就退化——这是所有工作共同的边界。

3. 核心机制：迁移范式如何层层演进

3.1 起点：把 egocentric 信号"蒸馏"进 exo 预训练

Ego-Exo 在标准 third-person 视频动作分类预训练之外，用离线 egocentric 模型为 Kinetics 视频生成三类伪标签——egocentricity 分数（Ego-Score）、被操纵物体（Object-Score）、手-物交互时空图（Interaction-Map）——以知识蒸馏损失注入视频模型。关键前置是不依赖配对或带动作标注的 ego 视频。结果在 Charades-Ego 上较 Kinetics 预训练基线高 +3.26 mAP，EPIC-Kitchens-100 达当时 SOTA。这个设定定义了后续所有工作的共同问题：在 ego 数据稀缺/无标注约束下，如何把 exo 的知识搬过来。

3.2 路线一：从"方法"到"数据集与基准"

Ego-Exo 之后，最显著的一步是把视角迁移基础设施化。Ego-Exo4D（Grauman 领衔）发布 1286 小时、740 名参与者、8 领域的时间同步第一+第三人称视频，并首次把 ego-exo relation（对象对应 correspondence、视角翻译 translation）形式化为可评测基准任务。这意味着"跨视角迁移"从一个论文技巧变成了一组有标准指标、有 leaderboard 的公共任务。EgoExoLearn 更进一步，把设定推进到异步、异环境的 demo-following——个体观看 exo 演示后在佩戴 gaze 设备的情况下于不同环境复现程序性任务。EgoExo-Fitness 则把双视角延伸到全身健身动作与"做得多好"的可解释动作评判。三者合起来，把 Ego-Exo 的"迁移"补全为"what / when / how well"三个维度的可评测问题。

3.3 路线二：桥接信号的迁移谱系

最实质的方法演进集中在"用什么桥接"。这一谱系清晰地显示了一条信号从强到弱、从前置苛刻到宽松的路线：

  • 时间同步：Synchronization is All You Need 证明，仅用带标注 exo 视频 + 无标注的同步 exo-ego 视频对做知识蒸馏，就能把时序动作分割模型从 exo 迁移到 ego，edit 分数 28.59，与监督式 ego-oracle（26.42）相当，远超仅用 exo 数据的基线（12.60）。同步是强信号，但需要采集同步对。
  • 无配对时序对齐：AE2（Xue & Grauman）把前置进一步放松到 unpaired：用 DTW 时序对齐作自监督目标、object-centric 编码器关注手与主动物体、反转帧作负样本，从无配对 ego-exo 视频学到细粒度视角不变表征（Break Eggs 动作相位分类 F1 66.23，超过最佳基线 GTA 的 56.86）。
  • 语言/伪配对：EgoInstructor 用 caption 语言对齐自动挖掘 ego-exo 对，EgoExoNCE 损失把跨视角视频特征对齐到共享文本语义，检索 third-person 教学视频作参考增强 egocentric 描述（InternVideo 在 CharadesEgo 上 Ego2Exo/Exo2Ego 分别提升 15.9%/7.7%）。
  • gaze：EgoExoLearn 的基准显示 gaze 是弥合视角差距的有效信号。

3.4 路线三：把"迁移"落成具体任务

在静态图像侧，一条独立的 affordance grounding 分支把 exo→ego 迁移做深做实。Learning Affordance Grounding from Exocentric Images 提出从第三人称人-物交互图像学可供性知识并迁移到 egocentric 图像（AGD20K，seen 设定 KLD 提升 17.2%），其期刊版 Grounded Affordance from Exocentric View 用 CFT 密集匹配显式建模"交互偏差"。LOCATE 把迁移细化到部件级，unseen 设定 KLD 提升 20.4%。ObjectRelator 则在 Ego-Exo4D 的对象对应任务上用多模态条件 + 跨视角对齐取得 SOTA（IoU Ego2Exo 39.7→44.3）。这条分支的共同教训是：迁移的精度取决于能不能把 exo 交互中"与任务相关的那部分"（部件、对象、可供性）单独拎出来，而不是全局搬特征。

4. 条件与分歧

  1. 桥接信号决定迁移上限，也决定适用边界。 同步、gaze、语言配对各有效，但都对应不同的采集成本与前置；没有哪种信号是普适的。EgoExoLearn 是这条边界的直接证据：即使 co-training ego+exo 并使用 gaze，跨视角关联准确率最高也只到约 55.3%（Exo2Ego test），异步跨视角桥接仍是显著未解决的 gap。
  2. 迁移精度受任务粒度制约。 全局特征蒸馏（Ego-Exo 原版）足够支撑动作识别，但 affordance 定位与对象对应需要部件级/对象级的显式对齐，说明"迁移"不是一个统一操作，而是随任务粒度的变化而分层。
  3. 证据面偏斜。 本池 20 条事件中绝大多数是 support，limit/conditional 证据稀缺，这更多是候选选择（引文派生的"有影响力后继"天然偏正向）所致，而非该领域没有失败；异步关联的 gap 是唯一显式负向信号。

5. 可操作框架

场景可用桥接信号参考工作
已有同步 ego-exo 视频对时间同步Synchronization / EgoExo-Fitness
只有无配对 ego+exo 视频时序对齐 / 伪配对AE2 / EgoInstructor
只有 exo 交互图像 + ego 物体图像部件/可供性定位Affordance / LOCATE
需评估动作质量专家评判 + gazeEgoExo-Fitness / EgoExoLearn

6. 研究空白与下一步

  • 异步跨视角的稳定桥接仍未解决（EgoExoLearn 的 55.3% 上限）——这是当前最明确的开放问题。
  • 现有工作多在单任务验证，缺乏跨任务统一的"视角不变表征"判据；AE2 提出的是时序对齐代理，EgoInstructor 提出的是语言对齐代理，二者如何统一尚无答案。
  • 引文池本身有边界：本 run 未覆盖关键词检索的全量候选，也未纳入 ego-only 的下游工作，因此"演进全貌"仍需一轮 planner 驱动的检索来补全。

7. 结论

Ego-Exo 的"第三人称→第一人称迁移"已从一个预训练技巧，长成了一个有数据、有基准、有任务谱系的研究范式。它留下的最有用的遗产不是某个蒸馏损失，而是一个判断：跨视角迁移的难点在桥接信号，不在迁移本身。谁能为"异步、无配对、弱监督"的真实场景找到更便宜、更稳的桥接信号，谁就掌握了这条线的下一程。

References

Accepted evidence

去重论文引用

  1. Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person PerspectivesKristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras · 2023
  2. Ego-Exo: Transferring Visual Representations from Third-person to First-person VideosYanghao Li, Tushar Nagarajan, Bo Xiong, Kristen Grauman · 2021
  3. EgoExo-Fitness: Towards Egocentric and Exocentric Full-Body Action UnderstandingYuan-Ming Li, Wei-Jin Huang, An-Lan Wang, Ling-an Zeng, Jingke Meng, Wei-Shi Zheng · 2024
  4. EgoExoLearn: A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real WorldYifei Huang, Guo Chen, Jilan Xu, Mingfang Zhang, Lijin Yang, Baoqi Pei · 2024
  5. EgoInstructor: Retrieval-Augmented Egocentric Video CaptioningJilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yue Zhang, Rui Feng · 2024
  6. Grounded Affordance from Exocentric ViewHongcheng Luo, Wei Zhai, Jing Zhang, Yang Cao, Dacheng Tao · 2022
  7. Learning Affordance Grounding from Exocentric ImagesHongcheng Luo, Wei Zhai, Jing Zhang, Yang Cao, Dacheng Tao · 2022
  8. Learning Fine-grained View-Invariant Representations from Unpaired Ego-Exo Videos via Temporal AlignmentZihui Xue, Kristen Grauman · 2023
  9. LOCATE: Localize and Transfer Object Parts for Weakly Supervised Affordance GroundingGen Li, Varun Jampani, Deqing Sun, Laura Sevilla-Lara · 2023
  10. ObjectRelator: Enabling Cross-View Object Relation Understanding Across Ego-Centric and Exo-Centric PerspectivesYu Fu, Runze Wang, Bin Ren, Guolei Sun, Biao Gong, Yanwei Fu · 2024
  11. Synchronization is All You Need: Exocentric-to-Egocentric Transfer for Temporal Action Segmentation with Unlabeled Synchronized Video PairsCamillo Quattrocchi, Antonino Furnari, Daniele Di Mauro, Mario Valerio Giuffrida, Giovanni Maria Farinella · 2023
完整证据附录（20 条正式证据事件）
  • Time range: 2026-02-13..2026-08-13
  • Events: 20
  • 每个事件一节,标题即锚点;trace-map 中的 event 链接跳转到这里。

EA-EGOEXO-2026-0001

  • Claim: Ego-Exo 从大规模第三人称视频中挖掘 Ego-Score、Object-Score、Interaction-Map 三类 egocentric 信号，并以知识蒸馏损失在预训练时注入视频模型，从而无需配对或动作标注的 ego 视频。
  • Stance: support | Confidence: direct
  • Paper: 2104.07905 Ego-Exo: Transferring Visual Representations from Third-person to First-person Videos
  • Locator: Abstract (full-text section)
  • Evidence: 摘要与方法部分直接陈述该核心思想与三类辅助任务。
  • Quote: “pre-training egocentric video models using large-scale third-person video datasets. Learning from purely egocentric data is limited by low dataset scale and diversity, while using purely exocentric (third-person) data introduces a large domain mismatch. Our idea is to discover latent signals in third-person video that are”
  • Authors: yanghao-li; tushar-nagarajan; bo-xiong; et al.

EA-EGOEXO-2026-0002

  • Claim: Ego-Exo 在 Charades-Ego 上较 Kinetics 预训练基线提升 +3.26 mAP，并在 EPIC-Kitchens-100 取得当时 SOTA。
  • Stance: support | Confidence: direct
  • Paper: 2104.07905 Ego-Exo: Transferring Visual Representations from Third-person to First-person Videos
  • Locator: 1 Introduction
  • Evidence: 实验部分报告了相对标准预训练基线的提升幅度。
  • Quote: “paradigm—by +3.26 mAP, and outperforms methods that specifically aim to bridge the domain gap between viewpoints. Finally, our pre-trained model achieves state-of-the-art results on EPIC-Kitchens-100 [ 11 ] , the largest available first-person dataset.”
  • Authors: yanghao-li; tushar-nagarajan; bo-xiong; et al.

EA-EGOEXO-2026-0007

  • Claim: 该工作提出从 exocentric 视角的 affordance grounding：从第三人称人-物交互图像学习可供性知识并迁移到 egocentric 物体图像，仅用 affordance 标签监督，并构建 AGD20K 数据集。
  • Stance: support | Confidence: direct
  • Paper: 2203.09905 Learning Affordance Grounding from Exocentric Images
  • Locator: Abstract (full-text section)
  • Evidence: 摘要直接陈述该任务设定与数据集。
  • Quote: “affordance grounding from exocentric view, i.e. , given exocentric human-object interaction and egocentric object images, learning the affordance knowledge of the object and transferring it to the egocentric image using only the affordance label as supervision. To this end, we devise”
  • Authors: hongcheng-luo; wei-zhai; jing-zhang; et al.

EA-EGOEXO-2026-0008

  • Claim: 在 seen 设定下，该方法 KLD 较最佳 saliency 模型提升 17.2%，较 affordance grounding 模型提升 13.3%。
  • Stance: support | Confidence: direct
  • Paper: 2203.09905 Learning Affordance Grounding from Exocentric Images
  • Locator: 5.2 Quantitative and Qualitative Comparisons
  • Evidence: 实验部分报告相对各基线的提升幅度。
  • Quote: “improves 17.2% compared to the best saliency model, 16.5% over the best weakly supervised object localization (WSOL) model, and 13.3% over the affordance grounding model in the “Seen” setting. Our method with the “Unseen” setting improves 10.2% compared to”
  • Authors: hongcheng-luo; wei-zhai; jing-zhang; et al.

EA-EGOEXO-2026-0016

  • Claim: 该期刊版在会议版基础上引入 Cross-view Feature Transfer（CFT）模块，通过密集匹配显式建模人与人之间的 interaction bias，扩展 AGD20K 至 26,117 图像。
  • Stance: support | Confidence: direct
  • Paper: 2208.13196 Grounded Affordance from Exocentric View
  • Locator: Abstract (full-text section)
  • Evidence: 引言与摘要陈述 CFT 模块与 interaction bias 建模。
  • Quote: ““interaction bias” between personas, mainly regarding different regions and views. To this end, we devise a cross-view affordance knowledge transfer framework that extracts affordance-specific features from exocentric interactions and transfers them to the egocentric”
  • Authors: hongcheng-luo; wei-zhai; jing-zhang; et al.

EA-EGOEXO-2026-0009

  • Claim: LOCATE 提出部件级知识迁移：定位 exo 交互区域、用 PartSelect 选出物体部件原型，再以此监督 egocentric affordance grounding，仅用图像级标签弱监督。
  • Stance: support | Confidence: direct
  • Paper: 2303.09665 LOCATE: Localize and Transfer Object Parts for Weakly Supervised Affordance Grounding
  • Locator: Abstract (full-text section)
  • Evidence: 摘要与方法部分直接陈述部件级迁移机制。
  • Quote: “transfer knowledge from images where an object is being used (exocentric images used for learning), to images where the object is inactive (egocentric ones used to test). To this end, we first find interaction areas and extract their feature embeddings. Then we learn to aggregate the embeddings into”
  • Authors: gen-li; varun-jampani; deqing-sun; et al.

EA-EGOEXO-2026-0010

  • Claim: 在 unseen 设定下，LOCATE 较 SOTA 方法 Cross-view-AG+ 提升 KLD 20.4%、SIM 33.3%、NSS 31.2%。
  • Stance: support | Confidence: direct
  • Paper: 2303.09665 LOCATE: Localize and Transfer Object Parts for Weakly Supervised Affordance Grounding
  • Locator: 4.2 Comparison to State-of-the-Art
  • Evidence: 实验部分报告相对 Cross-view-AG+ 的提升。
  • Quote: “improve the KLD by 20.4%, SIM by 33.3%, and NSS by 31.2% in the unseen setting. Cross-view-AG+ is an extended version of Cross-view-AG, but still performs the knowledge transfer based on global pooled embeddings at the image level, thus bringing only minor improvement.”
  • Authors: gen-li; varun-jampani; deqing-sun; et al.

EA-EGOEXO-2026-0011

  • Claim: AE2 从 unpaired 的 ego-exo 视频学习细粒度视角不变表征：用 object-centric 编码器关注手与主动物体，以 DTW 时序对齐作自监督目标，并用反转帧作负样本对比正则。
  • Stance: support | Confidence: direct
  • Paper: 2306.05526 Learning Fine-grained View-Invariant Representations from Unpaired Ego-Exo Videos via Temporal Alignment
  • Locator: 1 Introduction
  • Evidence: 摘要与方法部分直接陈述三个关键设计。
  • Quote: “unpaired data. In the unpaired setting, we know which human activity occurs in any given training sequence ( e.g. , pouring, breaking eggs), but they need not be collected simultaneously or in the same environment. The main idea of our self-supervised”
  • Authors: zihui-xue; kristen-grauman

EA-EGOEXO-2026-0012

  • Claim: AE2 在四个数据集、regular 与 cross-view 设定上全面超越 SOTA（如 Break Eggs 动作相位分类 F1 66.23 vs 最佳基线 GTA 56.86）。
  • Stance: support | Confidence: direct
  • Paper: 2306.05526 Learning Fine-grained View-Invariant Representations from Unpaired Ego-Exo Videos via Temporal Alignment
  • Locator: 5 Experiments
  • Evidence: 实验表 1 报告各数据集分类/检索指标。
  • Quote: “66.23 57.41 71.72 65.85 64.59 62.15 0.5109 0.6316 (B) Random Features 36.84 33.96 41.97 52.48 50.56 51.98 -0.0477 0.0050 ImageNet Features 41.59 39.93 45.52 54.09 27.31 43.21 -2.6681 0.0115 single-view TCN sermanet2018time 47.39 43.44 42.28 57.00”
  • Authors: zihui-xue; kristen-grauman

EA-EGOEXO-2026-0003

  • Claim: Ego-Exo4D 是迄今最大的时间同步第一+第三人称视频数据集，共 1286 小时、740 名参与者、覆盖 8 个领域技能活动，并配多模态与专家评论等语言标注。
  • Stance: support | Confidence: direct
  • Paper: 2311.18259 Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives
  • Locator: Abstract (full-text section)
  • Evidence: 摘要与数据集部分直接报告规模与模态。
  • Quote: “1,286 hours of video combined. The multimodal nature of the dataset is unprecedented: the video is accompanied by multichannel audio, eye gaze, 3D point clouds, camera poses, IMU, and multiple paired language descriptions— including a novel “expert commentary” done by coaches”
  • Authors: kristen-grauman; andrew-westbury; lorenzo-torresani; et al.

EA-EGOEXO-2026-0004

  • Claim: Ego-Exo4D 首次形式化 ego-exo relation 任务族（对象级对应 correspondence 与视角翻译 translation），使 first↔third-person 视角迁移成为可评测的基准任务。
  • Stance: support | Confidence: direct
  • Paper: 2311.18259 Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives
  • Locator: 1 Introduction
  • Evidence: 基准任务部分定义 ego-exo relation 家族。
  • Quote: “ego-exo relation , for relating the actions of a teacher (exo) to a learner (ego) by estimating semantic correspondences and translating viewpoints; 2. ego(-exo) recognition , for recognizing fine-grained keysteps and task structure; 3. ego(-exo) proficiency”
  • Authors: kristen-grauman; andrew-westbury; lorenzo-torresani; et al.

EA-EGOEXO-2026-0017

  • Claim: 该工作用无标注的同步 exo-ego 视频对做知识蒸馏，把时序动作分割模型从 exo 迁移到 ego，无需任何 ego 标签，并在 Assembly101 与 EgoExo4D 上验证。
  • Stance: support | Confidence: direct
  • Paper: 2312.02638 Synchronization is All You Need: Exocentric-to-Egocentric Transfer for Temporal Action Segmentation with Unlabeled Synchronized Video Pairs
  • Locator: Abstract (full-text section)
  • Evidence: 摘要与方法部分直接陈述该设定与方法。
  • Quote: “temporal action segmentation system initially designed for exocentric (fixed) cameras to an egocentric scenario, where wearable cameras capture video data. The conventional supervised approach requires the collection and labeling of a new set of egocentric videos to adapt”
  • Authors: camillo-quattrocchi; antonino-furnari; daniele-di-mauro; et al.

EA-EGOEXO-2026-0018

  • Claim: 同步蒸馏最佳模型在 Assembly101 上与监督式 ego-oracle 相当（edit 28.59 vs 26.42），远超仅用 exo 数据的基线（edit 12.60）。
  • Stance: support | Confidence: direct
  • Paper: 2312.02638 Synchronization is All You Need: Exocentric-to-Egocentric Transfer for Temporal Action Segmentation with Unlabeled Synchronized Video Pairs
  • Locator: 5.1 Performance of the Proposed Approach
  • Evidence: 实验表 1 报告 edit 分数对比。
  • Quote: “28.59 29.58 24.84 16.38 31.36 13 ✓ ✓ 28.13 28.75 24.16 15.79 32.67 14 Improvement w.r.t. baseline (line 10) Table 2 : Results for the v4 e4 view adaptation setting. Best results within groups are reported in bold. Best results among the adaptation”
  • Authors: camillo-quattrocchi; antonino-furnari; daniele-di-mauro; et al.

EA-EGOEXO-2026-0019

  • Claim: EgoInstructor 用自动挖掘的 ego-exo 视频对训练跨视角检索模块，检索语义相关的第三人称教学视频作参考，增强 egocentric 视频描述。
  • Stance: support | Confidence: direct
  • Paper: 2401.00789 EgoInstructor: Retrieval-Augmented Egocentric Video Captioning
  • Locator: Abstract (full-text section)
  • Evidence: 摘要直接陈述检索增强框架。
  • Quote: “retrieval-augmented multimodal captioning model that automatically retrieves semantically relevant third-person instructional videos to enhance the video captioning of egocentric videos, (2) for training the cross-view retrieval module, we devise an automatic”
  • Authors: jilan-xu; yifei-huang; junlin-hou; et al.

EA-EGOEXO-2026-0020

  • Claim: EgoInstructor 的跨视角检索模块在 7 个 benchmark 上一致提升，如 InternVideo 在 CharadesEgo 上 Ego2Exo/Exo2Ego 分别提升 15.9% 与 7.7%。
  • Stance: support | Confidence: direct
  • Paper: 2401.00789 EgoInstructor: Retrieval-Augmented Egocentric Video Captioning
  • Locator: 4.2.1 Results on Cross-view Retrieval
  • Evidence: 实验部分报告跨视角检索提升幅度。
  • Quote: “by 15.9% and 7.7% on CharadesEgo.”
  • Authors: jilan-xu; yifei-huang; junlin-hou; et al.

EA-EGOEXO-2026-0005

  • Claim: EgoExoLearn 将 ego-exo 迁移推进到异步、异环境的演示跟随（demo-following）设定：个体观看 exo 演示后在佩戴 gaze 设备的情况下于不同环境复现程序性任务，共 120 小时。
  • Stance: support | Confidence: direct
  • Paper: 2403.16182 EgoExoLearn: A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World
  • Locator: Abstract (full-text section)
  • Evidence: 摘要与数据收集部分直接描述该设定与规模。
  • Quote: “asynchronous procedural actions from different viewpoints. To this end, we present benchmarks such as cross-view association, cross-view action planning, and cross-view referenced skill assessment, along with detailed analysis. We expect EgoExoLearn can”
  • Authors: yifei-huang; guo-chen; jilan-xu; et al.

EA-EGOEXO-2026-0015

  • Claim: EgoExo-Fitness 提供 32 小时、1276 段同步 egocentric+exocentric 全身健身动作数据集，并引入技术关键点核验、自然语言评语、质量分等可解释动作评判标注。
  • Stance: support | Confidence: direct
  • Paper: 2406.08877 EgoExo-Fitness: Towards Egocentric and Exocentric Full-Body Action Understanding
  • Locator: 1 Introduction
  • Evidence: 摘要直接陈述数据集规模与标注。
  • Quote: “synchronized egocentric and exocentric videos of fitness activities to support future work on egocentric full-body action understanding. (b) EgoExo-Fitness provides abundant annotations, including two-level temporal boundaries and interpretable action judgement. (c) We benchmark”
  • Authors: yuan-ming-li; wei-jin-huang; an-lan-wang; et al.

EA-EGOEXO-2026-0013

  • Claim: ObjectRelator 针对 ego-exo 对象对应任务，用 MCFuse 融合文本描述与视觉掩码、XObjAlign 做自监督跨视角对象对齐，在 Ego-Exo4D 上取得 SOTA。
  • Stance: support | Confidence: direct
  • Paper: 2411.19083 ObjectRelator: Enabling Cross-View Object Relation Understanding Across Ego-Centric and Exo-Centric Perspectives
  • Locator: Abstract (full-text section)
  • Evidence: 摘要与引言直接陈述方法与任务。
  • Quote: “Bridging the gap between ego-centric and exo-centric views has been a long-standing question in computer vision. In this paper, we focus on the emerging Ego-Exo object correspondence task, which aims to understand object relations across ego-exo perspectives through segmentation. While numerous segmentation”
  • Authors: yu-fu; runze-wang; bin-ren; et al.

EA-EGOEXO-2026-0014

  • Claim: ObjectRelator 在 Ego-Exo4D 对象对应任务上较 PSALM 提升 IoU：Ego2Exo 39.7→44.3，Exo2Ego 44.1→49.2。
  • Stance: support | Confidence: direct
  • Paper: 2411.19083 ObjectRelator: Enabling Cross-View Object Relation Understanding Across Ego-Centric and Exo-Centric Perspectives
  • Locator: 4.1 Main Results on Ego-Exo4D
  • Evidence: 主结果表 2 报告 IoU 对比。
  • Quote: “improves the PSALM from 39.7 to 44.3 on Ego2Exo and from 44.1 to 49.2 on Exo2Ego. Regarding LE, CA, and VA, Our ObjectRelator consistently outperforms PSALM, achieving clear margins over all other competitors. These results well demonstrate the effectiveness of our approach. In”
  • Authors: yu-fu; runze-wang; bin-ren; et al.

EA-EGOEXO-2026-0006

  • Claim: EgoExoLearn 的 cross-view association 基准显示，即使 co-training ego+exo 并使用 gaze，模型跨视角关联能力仍有限（Exo2Ego test 最高约 55.3% top-1），暴露出桥接异步 ego-exo 活动的显著 gap。
  • Stance: gap | Confidence: direct
  • Paper: 2403.16182 EgoExoLearn: A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World
  • Locator: 4.2.1 Cross-view association
  • Evidence: 实验结果部分报告该基准的准确率并据此指出模型局限。
  • Quote: “55.3 51.1 Table 3 : Association accuracy in the cross-view association benchmark. In the fine-tuned setting, we adopt three kinds of data sources for training, i.e. , ego-only, exo-only, and hybrid ego-exo data. By leveraging gaze information during”
  • Authors: yifei-huang; guo-chen; jilan-xu; 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。