I’ve run an experiment with FasterRCNN but later I discovered that the tool I used to convert Yolo to Kitti (fiftyone) created annotations containing negative numbers like (-1, -2, -3, -4) for the coordinates of the bounding boxes that lie at the images borders. This is likely coming from rounding a substraction (but I am surprised why the library would allow that).
Anyway, I trained a model with annotations containing such negative numbers. I would like to know how does the TAO training docker deal with these annotations.
- Does it force them to be 0?
- Or does it not take them in account at all (i.e. remove them from the ground truth dataset) during the training loop?
Thanks
3 posts - 2 participants