Event
PRISM Uses Physical Noise Disentanglement to Improve Depth Synthesis Realism
1 reports · 1 sources
Xiujian Liang et al. propose the PRISM framework, which disentangles monolithic depth noise into two modalities—sensor failure and measurement inaccuracy—and distills features from 3D vision foundation models as spatial-semantic priors for physically grounded depth noise synthesis and sim-to-real transfer. The paper claims that its benchmark achieves the highest fidelity in noisy depth synthesis. In real-world robot experiments, PRISM enables downstream tasks to reach an average success rate of 92.1%, and the paper claims a significant improvement over monolithic baselines.
Report timeline
- PRISM Achieves Realistic Depth Synthesis via Physical Noise Disentanglement, with 92.1% Average Success Rate on Real-Robot Tasks
The paper proposes the PRISM framework, which disentangles depth noise into two modalities—sensing invalidation and measurement inaccuracy—and distills features from 3D visual foundation models as spatial-semantic priors.
arXiv cs.RO ·