Abstract
Egocentric videos present unique challenges for 3D reconstruction due to rapid camera motion and frequent dynamic interactions. State-of-the-art static reconstruction systems, such as MapAnything, often degrade in these settings, suffering from catastrophic trajectory drift and "ghost" geometry caused by moving hands. We bridge this gap by proposing a robust pipeline that adapts static reconstruction backbones to long-form egocentric video. Our approach introduces a mask-aware reconstruction mechanism that explicitly suppresses dynamic foreground in the attention layers, preventing hand artifacts from contaminating the static map. Furthermore, we employ a chunked reconstruction strategy with pose-graph stitching to ensure global consistency and eliminate long-term drift. Experiments on HD-EPIC and indoor drone datasets demonstrate that our pipeline significantly improves absolute trajectory error and yields visually clean static geometry compared to naive baselines, effectively extending the capability of foundation models to dynamic first-person scenes.
Abstract (translated)
第一人称视角视频因相机快速运动和频繁的动态交互，给三维重建带来了独特挑战。最先进的静态重建系统（如MapAnything）在此类场景中常因灾难性轨迹漂移和移动双手产生的“残影”几何体而性能下降。我们通过提出一种鲁棒的流程来弥合这一差距，该流程将静态重建主干网络适配于长视频形式的第一人称视角视频。我们的方法引入了一种掩码感知重建机制，该机制在注意力层中显式抑制动态前景，防止手部伪影污染静态地图。此外，我们采用分块重建策略并结合位姿图拼接，以确保全局一致性并消除长期漂移。在HD-EPIC和室内无人机数据集上的实验表明，与朴素基线相比，我们的流程显著改善了绝对轨迹误差，并生成了视觉上更干净的静态几何体，有效将基础模型的能力扩展到了动态第一人称场景。
URL
https://arxiv.org/abs/2603.22450
