How should be prepared data if file with prediction bbox or ground truth box is empty? about object-detection-metrics

[Kirayue]

Hi, @rafaelpadilla, I think I have a similar question.
I am trying to combine lib to my own project to calculate mAP. I followed Sample 2, and try to add bbox to BoundingBoxes, what if there is an image with no ground truth (e,g, only background), but the model has predictions on it. There is no doubt that we add a detection bbox to BoundingBoxes instance, but how about the ground_truth? Should I create a bbox instance with xywh all equal to 0 and add to BoundingBoxes instance, or just ignore it because no ground truth on it (i.e. there is a filename only has detection bbox but no ground truth bbox in BoundingBoxes)

[rafaelpadilla]

Dear @JamesHames,

Let me see if I understood your doubt: You have images that have no detections and/or no ground truth bounding boxes? If this is your question, you should leave the files empty, simple as that.

Regards,
Rafael

[HamsterHuey]

@Kirayue - From looking at the code, I believe you do not need to insert any BoundingBox object for the ground-truth if the image is a background only image with no actual ground-truth objects in them. The way the code is structured, any detection for such a "background" image will not be matched to a corresponding ground-truth, and as a result, will correctly be assigned as a false positive.