All training was done using rect=True to group images of similar aspect ratios into batches and pad each batch to the minimum rectangle that fits, rather than always padding to a square.
We should probably also set rect to true during inference as well if we want to see the benefit of this though.
Need to change the config and test to see if anything breaks or if there is an increase in mAP
All training was done using
rect=Trueto group images of similar aspect ratios into batches and pad each batch to the minimum rectangle that fits, rather than always padding to a square.We should probably also set
rectto true during inference as well if we want to see the benefit of this though.Need to change the config and test to see if anything breaks or if there is an increase in mAP