Artic-O: End-to-End Articulated Object Reconstruction via Latent Geometry Learning
Abstract Overview
Artic-O is an end-to-end feed-forward method for articulated object reconstruction from sparse multi-state images. The model maps observations from two articulation states into a pretrained latent geometry space, where a frozen flow-matching decoder reconstructs complete point clouds, including occluded structure. It then combines image tokens, geometry latents, and decoder features to predict active-part segmentation and articulation parameters within the same pipeline. The paper evaluates the method mainly on PartNet-Mobility and also reports qualitative tests on real phone captures and an inference-time extension to multi-part objects.
Novelty
The main novelty is unifying complete geometry reconstruction, active-part segmentation, and articulation estimation in a single feed-forward latent-geometry framework, rather than using a staged pipeline with post hoc articulation recovery. The method is also distinctive in using a frozen pretrained flow-matching shape decoder as a geometry prior, together with image-grounded part memory, a geometry-to-articulation curriculum, and decoupled two-pass training.
Results
On PartNet-Mobility, Artic-O improves geometry over LARM, reducing Chamfer Distance from 0.019 to 0.017 and increasing F-score from 93.8% to 95.4%. It attains comparable or better joint estimation on most reported metrics, including 96.0% axis-angle success, 98.0% axis-origin success, and 93.6% motion-range success, while running in about 0.32 seconds per object versus roughly 9 minutes for LARM. The paper also reports qualitative real-world reconstructions and improved multi-part reconstruction metrics over LARM without retraining.
Key Points
- Artic-O reconstructs geometry, active-part masks, and articulation parameters directly from sparse two-state images in one end-to-end feed-forward pipeline.
- A frozen pretrained flow-matching decoder supplies a latent complete-shape prior, while image-grounded part reasoning links geometry reconstruction to motion-part understanding.
- Experiments show better reconstruction quality than LARM with far lower inference cost, and the paper includes ablations, qualitative real-world examples, and a multi-part inference extension.