We haven't release the test config yet. The submit version is trained on the full 'trainval.txt' and has less data augmentations as illustrated in the paper. Our result is only trained by 'Car' category for test set.
We will complete the code after paper acceptance.
I found the corresponding description in your previous paper：
What I understand is that when you train on train set and validate on val set, the data augmentations are: random crop, random flip and photometric distortion.
when you train on trainval set and submit to benchmark, the data augmentations are: random flip and photometric distortion?
Do I understand correctly?
If my understanding is correct, why training on trainval set not using the random crop data augmentation?
We haven't release the test config yet. The submit version is trained on the full 'trainval.txt' and has less data augmentations as illustrated in the paper. Our result is only trained by 'Car' category for test set. We will complete the code after paper acceptance.
Originally posted by @ZrrSkywalker in https://github.com/ZrrSkywalker/MonoDETR/issues/12#issuecomment-1159518899
I found the corresponding description in your previous paper：
What I understand is that when you train on train set and validate on val set, the data augmentations are: random crop, random flip and photometric distortion.
when you train on trainval set and submit to benchmark, the data augmentations are: random flip and photometric distortion?
Do I understand correctly?