Experiment on Pascal VOC

[JiaPiangLi]

I have tried to train DINO-4Scale-R50 on Pascal VOC Dataset and only load the pretrain model of resnet 50 from torchvision. The AP0.5 is 0.585 at epoch 30, which is much lower than the benchmark of VOC. Does that mean the scale of VOC is still too small for DINO training or I have missed some other issues? My configurations are as following. GPU: RTX 3090 Batchsize: 4 lr: 1e-4 lr_backbone: 1e-5