Count the object, not the frame.
Field video shows one object in many frames and again when the camera passes back, so counting detections over-counts. The task was to count each object once, with an evaluation the team could trust.
The code and data stay private; a synthetic pond shows the same steps.
Aug 2026 to nowResearchNow
AI Researcher (Contract)
Louis Dreyfus Company
- Traced over-counting to duplicate tracks, not detector error, which cancelled a planned retrain.
- Cut over-counting at the hardest site by about three quarters with a gated merge across camera passes.
- Thesis (CountOnce): testing whether a video count can give a range that is right as often as it claims.
- Hand-audited 37 detections, tested 8 ways to merge duplicate tracks across camera passes, and gave every count a ±2 range.
Mar to Aug 2026Internship
Applied AI Engineer Intern · Computer Vision
Louis Dreyfus Company
- Built a YOLO and BoT-SORT pipeline that counts each object once in field video, deployed on Azure.
- Led annotation across 7 datasets (10,000+ images) and re-split validation by camera and site to stop leakage.
- 01VideoField footage, many frames per object
- 02Detect3-class detector
- 03TrackBoT-SORT, merge across passes
- 04Count±2 range per count
