Reproduce “LEO-VL: Efficient Scene Representation for Scalable 3D Vision-Language Learning” 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
Jiangyong Huang, Xiaojian Ma, Xiongkun Linghu, Junchao He, Qing Li, Song-Chun Zhu, Yixin Chen, Baoxiong Jia and 1 more
LEO-VL reduces 3D scene input for vision-language tasks by lifting multi-view image features into 3D and condensing voxel features into a position-aware grid. SceneDPO adds scene-contrast post-training. On SQA3D, it scores 60.8 exact-match and 63.7 refined exact-match with 750 scene tokens, versus GPT4Scene’s 59.4 and 62.4 with 8,262; this supports a capability–input-cost trade-off on that benchmark, not superiority across metrics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.