Why your DA-FRCNN implementation uses multi-scale training trick?

[tmp12316]

Thanks for your work, but I recently noticed another question about the input image scale.

As far as I know, the input min scale should be 600 for FRCNN-based DAOD frameworks, as shown in https://github.com/krumo/Domain-Adaptive-Faster-RCNN-PyTorch/blob/df0488405a7679552bc2504b973e29178c141b26/configs/da_faster_rcnn/e2e_da_faster_rcnn_R_50_C4_cityscapes_to_foggy_cityscapes.yaml#L24

But It seems that AT uses multi-scale training in all configs?

adaptive_teacher/configs/Base-RCNN-C4.yaml

Line 17 in cba3c59

MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)

[yujheli]

Sry I did not copy the correct Base-RCNN-C4 I used internally. I copied the one from detectron2. Will update.

[yujheli]

Hi @tmp12316 , I tested with the correct config file without multi-scale trick and got 45.6 AP@50 on Clipart1k using batch size 4. Will update the experiment with more batch size once I have enough local training resources.

[tmp12316]

@yujheli

Hi, I think that this result seems to be much more reasonable. Thanks for your kind reply!