🌍 Spatial-SSRL: Spatial Reasoning with Vision-Language Models

This demo showcases the spatial reasoning capabilities of vision-language models trained to understand 3D spatial relationships from 2D images.

Features

• 3D Location Understanding: Determine which objects are closer or further from the camera
• Orientation Analysis: Understand which direction objects are facing
• Relative Positioning: Answer questions about object positions relative to each other
• Step-by-step Reasoning: The model provides detailed reasoning before answering

How to Use

1. Upload an image
2. Ask a question about spatial relationships in the image
3. The model will provide a detailed answer with reasoning

Example Questions

• "Which object is further away from the camera? A. boat B. fire hydrant"
• "Are the kid and the teddy bear facing same or similar directions?"
• "If I stand at the recreational vehicle's position facing where it is facing, is the dog in front of me or behind me?"

The model is trained to provide answers in a structured format with reasoning enclosed in <think> tags and final answers in \boxed{}.