Venue
Computer Vision and Pattern Recognition
Domain
computer vision, natural language processing, robotics
Understanding and reasoning about spatial relationships is a fundamental capability for Visual Question Answering (VQA) and robotics.While Vision Language Models (VLM) have demonstrated remarkable performance in certain VQA benchmarks, they still lack capabilities in 3D spatial reasoning, such as recognizing quantitative relationships of physical objects like distances or size difference.We hypothesize that VLMs' limited spatial reasoning capability is due to the lack of 3D spatial knowledge in training data and aim to solve this problem by training VLMs with Internet-scale spatial reasoning data.To this end, we present a system to facilitate this approach.We first develop an automatic 3D spatial VQA data generation framework that scales up to 2 billion VQA examples on 10 million real-world images.We then investigate various factors in training recipe including data quality, training pipeline and VLM architecture.Our work features the first Internet-scale 3D spatial reasoning dataset in metric space.By training a VLM on such data, we significantly enhance its ability on both qualitative and quantitative spatial VQA.Finally, we demonstrate that this VLM unlocks novel downstream applications in chain-of-thought spatial reasoning and robotics due to its quantitative estimation capability.
The paper introduces SpatialVLM, a new framework enhancing Vision-Language Models (VLMs) with spatial reasoning capabilities through training on Internet-scale spatial reasoning datasets. A key feature is the automatic generation of a 3D spatial Visual Question Answering (VQA) dataset comprising around 2 billion examples from 10 million real-world images. By combining various computer vision techniques, the authors create dense 3D annotations, allowing VLMs to better understand spatial relationships and perform quantitative estimations. The authors demonstrate improvements in spatial reasoning tasks, showing that VLMs can now assist in complex applications in robotics and can serve as reward annotators. The paper emphasizes the significance of quality training data and explores the relationship between model architecture and data quality, concluding that enhanced spatial reasoning capabilities can be trained into VLMs by addressing dataset limitations. The research presents promising findings for future applications in robotics and spatial reasoning.
This paper employs the following methods:
- SpatialVLM
- GPT-4V
- PaLM-E
- PaLI
The following datasets were used in this research:
- Quantitative Spatial VQA performance
- Qualitative Spatial VQA performance
- Accuracy
- Enhanced qualitative and quantitative spatial VQA capabilities in VLMs
- Successful real-world applications in chain-of-thought reasoning and robotics
The authors identified the following limitations:
- Number of GPUs: None specified
- GPU Type: None specified
vision-language models
spatial reasoning
3D understanding
visual question answering
robotics