Computer vision / Sep 2026
CLIP Semantic Image Search & Retrieval Evaluation
Vision-language search and retrieval evaluation
Searches a photographed image collection by text or example image and compares CLIP retrieval with a color-histogram baseline.
95.2%same-category top-1 · 21 leave-one-out queries
0.799mean average precision · same evaluation
Search and evaluation
The Gradio app ranks text-to-image results from cached CLIP embeddings and also supports image-to-image retrieval. An HSV histogram provides a color-only baseline.
Benchmark scope
The follow-up leave-one-out evaluation uses 21 query photos from three clearly labeled groups—mugs, shoes, and cans—against all 29 photos. Each query image is excluded from its results. Mean average precision uses same-category photos as relevant items. The collection is small, so these results describe this controlled test rather than general-world retrieval performance.
What I built
- Built text-to-image and image-to-image search with CLIP ViT-B/32, a Gradio interface, and cached 512-dimensional embeddings; encoded the 29-photo collection in 0.609 seconds on CUDA.
- In a 21-query leave-one-out evaluation against the full collection, achieved 95.2% same-category top-1 retrieval (20/21) versus 33.3% (7/21) for the HSV color-histogram baseline—a 61.9-percentage-point difference.
- Achieved 0.799 mean average precision versus 0.343 for the HSV baseline, while analyzing text-query failures involving exact object counts and ambiguous descriptions.