Reproduce “Q-GeoMem: Question-Guided Geometric Memory for Video Spatial Reasoning” 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
Xianqiang Gao, Qizhi Chen, Delin Qu, Haoming Song, Zhigang Wang, Bin Zhao, Dong Wang, Xuelong Li
Q-GEOMEM introduces a question-guided memory system for video spatial reasoning that separates dense short-term context from compact long-term evidence. By combining camera-conditioned geometry fusion with relevance–novelty-based memory selection, it reaches 71.2 on VSI-Bench and 67.8 on VSTI-Bench, outperforming the reported VLM² and VLM-3R baselines while using a fixed eight-entry evidence bank.
Key insights
- Memory capacity alone is not enough for spatial reasoning: selecting complementary evidence is more effective than retaining all sampled frames or using FIFO replacement.
- The relevance–novelty product avoids repeatedly storing similar views of the queried object while still filtering out diverse but question-irrelevant observations.
- Camera-aware short-term memory is particularly useful for local motion and viewpoint-dependent judgments, while the semantic-geometric bank contributes most to temporally distributed tasks such as appearance order and size estimation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.