DeepSurv
Drone video, understood as it arrives.
Detection, segmentation, and tracking on a live drone feed, logged as a dataset you can query rather than footage someone has to scrub.
Built with
2026 · AI Systems
What it does
YOLOv8-Seg multi-class detection with instance segmentation
FastSAM pass for out-of-vocabulary objects
ByteTrack persistent identity across frames
Structured JSON + Parquet logs, queryable after the run
WebSocket streaming API and reproducible dataset benchmarking
The long version
A drone feed is a hard case for perception — small objects, fast ego-motion, and a scene that never sits still. DeepSurv processes it end to end at interactive latency by layering models with different strengths rather than asking one model to do everything.
YOLOv8-Seg provides 80-class detection with instance masks, FastSAM fills in comprehensive scene segmentation for objects the class vocabulary does not cover, and ByteTrack keeps object identities stable across frames so a track means something over time.
Everything the system observes is written as structured JSON and Parquet, which is the difference between a video someone has to scrub through and a dataset that can be queried after the fact. A WebSocket API streams detections live to remote dashboards, and a built-in Stanford Drone Dataset loader keeps benchmark runs reproducible.