Open zhifanlight opened 4 years ago
We used the TTA setting mentioned in the paper, but we didn't use the TTA code here in mmdetection due to the speed and performance issues. Please refer to the TTA code of Detectron/Detectron2. We adopted their augmentation process. Thanks.
@joe-siyuan-qiao Hi, thanks for your great work. But what is the correct TTA settings to get 54.7mAP on coco val2017?
When using
img_scale=[(1200, 800), (1500, 1000), (1800, 1200), (2100, 1400), (2400, 1600)], flip=True
(same with paper), it takes 7 hours on 8x10180Ti, but only get 53.7mAP, compared with defaultimg_scale=(1333, 800), flip=False
, 52.9mAP.