Loading content…
Independent
Independent interactive visual lab for simulating multi-object trackers to discover their weaknesses using ground truth versus actual behavior.
PythonFastAPIREST APIsAsync ProcessingNext.jsTypeScriptDocker ComposeOpenCVUltralyticsByteTrackBoT-SORTOC-SORTDeepOCSORT
An independent project created to investigate and stress-test visual tracking architectures. In this project, I will be adding more tracker simulations to investigate and benchmark their weaknesses using ground truth versus actual behavior under stress.
It provides an interactive visual computer vision lab designed to discover why and how multi-object trackers fail. Features dual operating modes: a Synthetic Simulator with deterministic ground truth physics (linear velocity, crossing trajectories, occlusion, camera jitter, detection noise) and a Real Video YOLO Mode for running detectors and trackers on live footage. Implements side-by-side MOTA, MOTP, IDF1, precision/recall, and ID switch evaluations alongside automated natural-language weakness report cards across 16 tracker algorithms.
Multi-object trackers frequently fail under severe occlusion, sensor noise, crossing trajectories, and missed detections, but diagnosing failure root causes in production video feeds is difficult without comparing actual behavior against deterministic ground truth.
Built a dual-mode evaluation framework combining an interactive Next.js and TypeScript frontend with a high-performance Python/FastAPI backend utilizing REST APIs and async processing, orchestrated via Docker Compose. Generates parametric synthetic ground truth and integrates real-world YOLO inference with 16 tracker algorithms (ByteTrack, BoT-SORT, OC-SORT, DeepOCSORT, OpenCV classic trackers), computing real-time MOT metrics and weakness report cards.
Enables reproducible side-by-side benchmark evaluation of 16 tracking algorithms, discovering weaknesses and surfacing identity switches, false trajectories, and localization drift under adversarial conditions.
Skills
