These are the scripts used in How to Repair Object Detectors? A Benchmark, an Empirical Study, and a Way Forward.
https://github.com/YYue000/mmdetection/tree/repair
https://github.com/YYue000/Neighbor2Neighbor
Corruptions are available in imagecorruptions[https://github.com/bethgelab/imagecorruptions]. Models are based on mmdetection[https://github.com/YYue000/mmdetection/tree/repair].
Failure samples are collected with scripts in benchmark/. We first compute image-wise AP with benchmark/setup/eval_one_img.py. Then we setup the failure dataset with benchmark/setup/setup.py
Analysis for failure samples are conducted with scripts in benchmark/post.
To automatically setup the benchmark with many experiments, please refer to repair_benchmark/post/setup.py.
Directly applied with mmdetection. Scripts are availble in repair_benchmark/*.sh.
We use a refactored version of Neighbor2Neighbor[https://github.com/YYue000/Neighbor2Neighbor].
Please refer to imagenet_bn_calibration/cal.py.