sdmac/Urban-Draw-Detection

UrbanDraw AI

This Space detects four broad themes in children's drawings of urban environments and returns:

• —an annotated drawing;
• —normalized bounding boxes, confidence values, spatial zones, and category counts;
• —overall and detected-region HSV color summaries;
• —a cautious, non-clinical interpretation of at most 100 words.

Model

The detector is a YOLOv9e model trained at 640-pixel resolution for the four project categories.

The reproducible training notebook is available in the companion GitHub repository. The research dataset is not distributed with this Space.

The application loads the open Qwen/Qwen2.5-0.5B-Instruct language model locally. Set the Space variable USE_LOCAL_LLM=0 to use the deterministic interpretation fallback only. You can select another compatible model with the LLM_MODEL_ID Space variable.

Website API

The public Gradio endpoint is named /analyze. The website can call it through @gradio/client without exposing a Roboflow or OpenAI key.

Color analysis runs inside this Space, so the website does not need to duplicate the HSV logic. It can embed the Space or present the /analyze response in a custom interface.

Responsible-use scope

Outputs are drawing-level exploratory observations for participatory urban research. They must not be used to diagnose a child, infer trauma or mental-health conditions, or make individual decisions. Color and spatial associations are culturally and contextually dependent. Bounding-box colors may include neighboring background because the dataset contains boxes rather than segmentation masks. The dataset citation and DOI will be added after publication.