Self-Supervised AI Framework FoundObj Enables Label-Free 3D Object Segmentation
Researchers have introduced FoundObj, a novel framework for 3D object segmentation in complex scene point clouds that operates without any scene-level human annotations during training. Existing methods are limited to identifying simple objects due to insufficient object priors. FoundObj employs a superpoint-based object discovery agent that incrementally merges neighboring superpoints, guided by semantic and geometric reward modules. These modules leverage priors from self-supervised 2D/3D foundation models, providing complementary feedback through reinforcement learning. The approach consistently outperforms existing baselines across diverse benchmarks, demonstrating robust identification of multi-class objects. The paper is published on arXiv with ID 2605.27178.
Key facts
- FoundObj is a framework for 3D object segmentation without scene-level human annotations.
- It uses a superpoint-based object discovery agent that merges neighboring superpoints.
- Semantic and geometric reward modules guide the agent using self-supervised 2D/3D foundation models.
- Reinforcement learning enables robust multi-class object identification.
- The method outperforms existing baselines on diverse benchmarks.
- The paper is available on arXiv (ID: 2605.27178).
Entities
Institutions
- arXiv