Project Overview
Objective
Built text-to-video semantic scene retrieval with multilingual query processing.
Stack
OpenCVCLIP (ViT-B/32)FAISSFastAPIReact.jsTailwind CSS
Delivery highlights
- Built a prototype that samples video frames with OpenCV, encodes the keyframes and natural-language queries with CLIP (ViT-B/32), and stores the precomputed frame embeddings in FAISS for semantic scene retrieval.
- Added Thai-to-English query translation before embedding so multilingual searches can be compared in CLIP's shared vector space. Search results return relevant video segments with similarity scores and timestamps rather than treating the video as one undifferentiated file.
- Exposed indexing and search through FastAPI and provided a React interface for entering queries, reviewing ranked scenes, and navigating to matching moments. The audited source is a development snapshot; authentication, durable job processing, and production deployment were not confirmed.