All the things you need is in PythonAPI/

sudo pip3 install Cython
cd PythonAPI
make

cd PythonAPI
python3 eval.py <path to GT json> <path to predictions json> 

The JSON files for Ground Truth (GT) and Predictions are expected to be in COCO format (ignore the segmentation stuff), thought not all are essential (for example things like licenses need not to be included in the json)

GT JSON should have images, categories and annotations.

Prediction JSON can just be a list of predicted bounding boxes with image_id that corresponds to the images given in GT JSON

[
  {"image_id": 0, "category_id": 1, "bbox": [0.0, 193.94281005859375, 131.76132202148438, 79.2210693359375], "score": 0.8640222549438477},
  {"image_id": 1, "category_id": 1, "bbox": [131.12359619140625, 174.27325439453125, 309.98846435546875, 62.258544921875], "score": 0.436798095703125},
  ...
]

• PythonAPI/asym_eval.py: Use this when only want to evaluate a subset of the images that are predicted on. Create a GT json that only contains the subset you want (important point: make sure the image_ids are unchanged), then test it against the prediction json (prediction json can contain images not in the subset GT json).

• PythonAPI/visualise_bb.py: Outputs visualised GT and predicted bbs on each validation image in the GT json. Able to take in asym pred vs GT (similar to asym_eval.py)