Computer Vision Competition
Neural Debris Removal Competition
A computer-vision competition project using a supplied RetinaNet detector. I compared unlearning and detection-filtering approaches, checked what each change removed, and reviewed how much normal detection behavior it preserved.
Period
Jul 2026
Tools / Tech
PythonPyTorchRetinaNetScikit-learnNumPyPytest
Why I built it
Removing an unwanted detection is only half the problem: an aggressive change can also erase useful results. I wanted to compare both effects before choosing a submission.
Links
What it includes
- Reproduces the supplied model's inference and compares it with reference predictions.
- Compares training, distillation, and object-level filtering experiments using held-out examples.
- Records experiment settings, validates submission files, and checks that repeated runs produce the same output.
What I worked on
- Chose the experiment order and reviewed the trade-off between removing unwanted detections and preserving normal ones.
- Used AI assistance for implementation, then inspected predictions and failure cases to choose follow-up experiments.
- The July 2026 experiment log records a public-leaderboard score of 222.8139 and rank 41 at the time of submission.
What I Learned
- Learned how a local evaluation can miss behavior that appears on new examples.
- Practiced comparing intermediate predictions and keeping experiment records useful for later review.