OccluCheck

OccluCheck is a desktop + web app that stress-tests a robot’s vision pipeline for occlusion and lighting failures using your existing camera feeds or recorded ROS bags. You import a short set of representative scenes, then the app automatically generates controlled “what-if” variations: partial occlusions, glare, motion blur, low light, dirty lens, and sensor noise. It runs your detector/segmenter (ONNX/TensorRT) across these variants and produces a failure report: where confidence collapses, which classes are most brittle, and which camera angles are risky. The output is practical: a ranked list of scenarios to re-capture, suggested camera repositioning heuristics, and a regression suite you can run on every model update. This is not a magic model builder; it’s a reliability harness that helps teams stop shipping vision systems that break in the real world.