SpatialBot: Precise Spatial Understanding with Vision Language Models

TL;DR

SpatialBot enhances vision language models' spatial understanding by integrating RGB and depth data, supported by a new dataset and benchmark, leading to significant performance improvements in spatial reasoning tasks.

Contribution

Introduction of SpatialBot with depth integration, creation of SpatialQA dataset for depth-related questions, and development of SpatialBench for comprehensive spatial understanding evaluation.

Findings

SpatialBot outperforms existing models on spatial understanding benchmarks.

Training on SpatialQA improves depth-related reasoning capabilities.

Model shows significant gains in Embodied AI tasks.

Abstract

Vision Language Models (VLMs) have achieved impressive performance in 2D image understanding, however they are still struggling with spatial understanding which is the foundation of Embodied AI. In this paper, we propose SpatialBot for better spatial understanding by feeding both RGB and depth images. Additionally, we have constructed the SpatialQA dataset, which involves multi-level depth-related questions to train VLMs for depth understanding. Finally, we present SpatialBench to comprehensively evaluate VLMs' capabilities in spatial understanding at different levels. Extensive experiments on our spatial-understanding benchmark, general VLM benchmarks and Embodied AI tasks, demonstrate the remarkable improvements of SpatialBot trained on SpatialQA. The model, code and data are available at https://github.com/BAAI-DCAI/SpatialBot.