Disclaimer
This is a personal learning project and local MVP. It is not production software, and the implementation may change as I experiment with video indexing, image embeddings, search quality, and UI workflows.
Video Search by Screenshot indexes a local video folder and lets you search the indexed videos with a screenshot. The current implementation is designed for a curated video subset, local cache storage, and iterative experimentation.
uv sync --python C:\Python311\python.exeThe project uses uv for Python environment and command execution.
uv run python -m app.env_check
uv run python -m app.benchmark_embed --device cuda --batch-size 16
uv run python -m app.index_videos --input "Z:\SampleVideos" --fps 1
uv run python -m app.query_image --image ".\query.jpg" --top-k 10
uv run python -m app.eval --queries ".\examples\sample_queries.jsonl"uv run python -m app.guiThe GUI starts a local browser interface. It supports choosing a video folder and cache root with native directory pickers, indexing frames into the local cache, entering or uploading a screenshot query, and browsing grouped matches with thumbnails, timestamps, scores, and source paths.
Indexing runs are persisted under the selected cache root in index_runs.json.
The GUI shows resumable runs on startup, checks whether the source folder still
exists, and can pause or resume indexing between videos.
.\scripts\benchmark_index_windows.ps1 -Input "Z:\SampleVideos" -Device cuda -BatchSize 32./scripts/benchmark_index_macos.sh --input "/Volumes/NAS/SampleVideos" --device auto --batch-size 32