Output detection metrics/bounding boxes overlaid on ground truth boxes

Using the tutorial https://gilberttanner.com/blog/detectron-2-object-detection-with-pytorch, I managed to get my custom dataset running and detecting (NOTE: My dataset is a single class, and I am doing plain old non-rotated bounding boxes). I have registered my coco instances using the following chunk of code to get separate train/validation sets:

# Register training dataset register_coco_instances("breaks_train", {}, "/gdrive/blah/drive/here/folder/train.json", "/gdrive/some/folder/here/trainandvalidfolders/Train")

# Register validation dataset register_coco_instances("breaks_valid", {}, "/gdrive/blah/drive/here/folder/valid.json", "/gdrive/some/folder/here/trainandvalidfolders/Valid")

After training, I would like to get back the AP and AR, and any other metrics that may be useful. Ideally, I’d like to be able to ouput a Precision-Recall curve for various IOU values like in the COCO challenge. I’m kind of a noob to object detection and pytorch in general, so any code snippets or chunks would be very appreciated!

[Labaien96]

Hi @djpecot , I also want to visualize bot GT boxes and predicted boxes in the same image, do you know how to do it?

[ppwwyyxx]

No. You can use https://detectron2.readthedocs.io/modules/utils.html#module-detectron2.utils.visualizer to draw them yourself.