Closed avecplezir closed 2 years ago
The batch dim is indeed flattened, but the following line should allow matching a source label only to a target in the corresponding image, specifically look at C.split
:
https://github.com/amirbar/DETReg/blob/4f9cb9aac896110562b0a36c2c932fb25b5e4a54/models/matcher.py#L95
Hi! Thank you for your code and work!
It seems that it is possible to match predictions of image A with targets of image B if the batch is bigger than one. Does it lead to the potentially worse solution when finetuning on COCO and pertaining?
https://github.com/amirbar/DETReg/blob/4f9cb9aac896110562b0a36c2c932fb25b5e4a54/models/matcher.py#L73