Visual Question Answering With Spatial Reasoning

Tiny vision-language model for edge devices.

Unique: Implements region encoding subsystem that maps pixel-level coordinates to semantic embeddings, enabling spatial reasoning without post-hoc bounding box detection; uses transformer cross-attention between vision and text embeddings to ground language generation in visual features, avoiding separate vision-text alignment modules.

vs others: Faster and more memory-efficient than BLIP-2 or LLaVA for VQA tasks due to smaller parameter count; maintains spatial reasoning capabilities that pure image captioning models lack.