This is the code for the paper
Learning non-maximum suppression. Jan Hosang, Rodrigo Benenson, Bernt Schiele. CVPR 2017.

You can find the project page with downloads here: https://mpi-inf.mpg.de/learning-nms

We have done some optimization on the original author's code (in the production of manual features, we have compressed the size of multi category features). This optimization has greatly improved the multiclass AP while keeping the map from getting worse .

The following is the validation data after the 300000 training round. (tensorflow)

The following is the validation data after the 20000 training round. (pytorch)

• torch : 1.7.1.post2
• numpy : 1.19.5
• protoc : libprotoc 3.13.0
• scipy : 1.1.0

Run make in the home directory to compile protobufs

Download the files according to the addresses in ./data/README and ./data/coco/annotations/README files

run python train.py --config=experiments/coco_person/conf.yaml