Q-GeoMem: A Question-Guided Geometric Memory Framework Revolutionizing Video Spatial Reasoning

Q-GeoMem is an innovative framework for video spatial reasoning. Its core lies in integrating camera-conditioned geometry into visual tokens via a question-guided geometric memory mechanism, addressing the issues of memory redundancy and weak long-range reasoning in traditional methods, and achieving state-of-the-art performance on VSI-Bench and VSTI-Bench.

Video spatial reasoning needs to handle time accumulation, perspective changes, geometric perception, and question relevance requirements (e.g., robots answering the positions of objects in the environment). Existing models treat memory as a general temporal cache, which easily introduces redundant geometric information and weakens long-range reasoning capabilities.

Q-GeoMem consists of three core components:

1. Camera-conditioned geometry injection: Encode camera parameters and inject them into visual tokens to explicitly understand spatial positions;
2. Dual memory system: A fine-grained context store holds recent features and camera states, while a semantic-geometric evidence store holds long-range compressed evidence;
3. Evidence scoring mechanism: Filter frames based on question relevance (Q-Former embedding) and novelty scores; memory updates use a capacity replacement rule.

Q-GeoMem achieves state-of-the-art results on VSI-Bench and VSTI-Bench, with outstanding performance in long-range reasoning and camera movement scenarios. Ablation experiments demonstrate:

• Removing the scoring mechanism leads to performance degradation;
• The dual memory system is complementary and optimal;
• Camera-conditioned injection is critical in perspective change scenarios.

Traditional methods (uniform sampling, spatial-unconstrained attention, fixed capacity) do not consider the varying needs of different questions. Q-GeoMem constructs dynamic memory for different questions through question embedding, relevance scoring, and adaptive memory, improving reasoning efficiency and accuracy.

Q-GeoMem can be applied in embodied intelligence (robot environment understanding), autonomous driving (3D structure cognition), augmented reality (perspective spatial relationships), video question answering (spatial reasoning tasks), and other fields.

Current limitations: High computational overhead, fixed memory capacity, unvalidated generalization, insufficient multimodal expansion. Future directions: Lightweight architecture, unsupervised learning, expansion to complex scenarios, integrating world models for predictive reasoning.