data augmentation in inference ?

[ghost]

Hi, I notice there are some data augmentation procedures during the inference stage since the logger shows:

[DatasetMapper] Augmentations used in inference: [ResizeShortestEdge(short_edge_length=(800, 800), max_size=1333, sample_style='choice')]

I wonder whether we can turn off the data augmentation procedure during the inference stage ? Does this harm the model performance ? Besides, after data augmentation, the boxes in the original images will mismatch with the golden ones, how do you solve this problem ? Thanks.

[ShoufaChen]

Hi,

Thanks for your interest.

This data augmentation at the inference state is borrowed from Sparse R-CNN, Deformable DETR, and DETR.

You can turn off this data augmentation to test its performance. I think it is not complicated since the evaluation didn't cost much computational resources.

When conducting data augmentation, the corresponding box annotations will also be adjusted. Therefore, there isn't a mismatching problem.

[tzayuan]

Hi, I want to eval the model by python train_net.py --num-gpus 1 --config-file configs/diffdet.coco.res50.300boxes.yaml --eval-only MODEL.WEIGHTS diffdet_coco_res50_300boxes.pth, meanwhile, I want to turn off above data augmentation ([DatasetMapper] Augmentations used in inference: [ResizeShortestEdge(short_edge_length=(800, 800), max_size=1333, sample_style='choice')]), but I can't turn off the data augmentation. I want to detect the image of its original resolution rather than resized resolution. If you could give me some suggestions? @ShoufaChen Thanks!