Reproduce “Objectron: A Large Scale Dataset of Object-Centric Videos in the Wild with Pose Annotations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Adel Ahmadyan, Liangkai Zhang, Jianing Wei, Artsiom Ablavatski, Matthias Grundmann
Objectron introduces a large-scale dataset of 14,819 object-centric videos comprising 4 million images, annotated with precise, metric-scale 3D bounding boxes. This resource aims to advance 3D object understanding by providing diverse real-world data and supporting tasks like 3D object detection and tracking.
Key insights
- Object-centric video collection with AR metadata allows for efficient and scalable annotation of metric-scale 3D poses in diverse real-world settings.
- The dataset's temporal consistency and multi-view information are crucial for advancing tasks beyond static 3D detection, such as 3D tracking and shape reconstruction.
- Baseline models, despite being lightweight and mobile-optimized, achieve measurable performance, highlighting the dataset's utility and the remaining challenges, especially for symmetric objects and those with varying textures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.