View Project Report | View Code
Overview
LensDB is a compressed learned index for surveillance and traffic video that decouples ingestion from interactive querying. Instead of running object detection per frame, it stores a sparse set of representative frames as CLIP embeddings and answers exploratory queries directly in the latent space, avoiding repeated video decode from network-attached storage.
Key Achievements
- Up to 99.991% storage reduction with sub-second query latency vs. exhaustive YOLO baselines.
- Compresses a 1.4 GB video into ~1.5 MB of embeddings while supporting approximate car-count queries.
- Keyframe filtering removes 90–99% redundant frames before embedding using FrameDiff, SSIM, MOG2, and optical flow.
Technical Implementation
- Ingestion: 1 FPS sampling → heuristic keyframe selection → CLIP (ViT-B/32) image embeddings → FAISS index + timestamp metadata map.
- Query: CLIP text encoder retrieves top-k via FAISS, then an MLP predicts object count for threshold filters (Count ≥ T), with temporal expansion via metadata.
Results
- Achieves F1 = 0.963 for event detection at T ≥ 1, with expected precision trade-offs at higher count thresholds.