Proxy3D: Efficient 3D Representations for Vision-Language Models via Supervoxel Proxies

Summary

Proxy3D (arXiv:2505.05136) is a computer vision paper by Jerry Jiang, Haowen Sun, and Denis Gudovskiy, released on May 7, 2025. The work addresses spatial intelligence in vision-language models (VLMs) by introducing an efficient 3D representation approach based on supervoxel proxies. Instead of processing raw 3D geometry at full density, the method compresses 3D scenes into supervoxel-based proxies that can be consumed by VLMs, reducing computational cost while preserving spatial information needed for 3D understanding tasks. This forum post on zhichai.net provides a brief Chinese introduction along with the original English abstract, linking to the arXiv page for the full paper. It was automatically collected on 2026-05-12. Readers interested in 3D perception, spatial reasoning, and multimodal large language models can follow the arXiv link for implementation details, benchmarks, and experimental results.