search

chenqi1126 / SIPE

[CVPR 2022] Self-supervised Image-specific Prototype Exploration for Weakly Supervised Semantic Segmentation
MIT License
73 stars 10 forks source link

About the pertained backbone for DeepLab V2 #11

Closed tomtomsa closed 1 year ago

tomtomsa commented 1 year ago

Hi, Thanks for your interesting work. I notice that you have trained a DeepLabV2 semantic segmentation network. Regarding the pre training of ResNet101, is it on COCO or ImageNet? I notice that you give two repo (deeplab-pytorch and seamv1.) as guidance, but deeplab-pytorch uses COCO pretrained weights, and seamv1 uses ImageNet pretrained weights. Since you didnot explain it in the paper or this repo, I open this issue for the answer. Looking forward to your reply.

Sincerely, Tomsa.

chenqi1126 commented 1 year ago

@tomtomsa,

Thanks for your attention!

Yes - for a fair comparison on VOC dataset, we follow the previous works that use COCO pretrained weights to initialize ResNet101 and ImageNet pretrained weights to initialize WideResNet38. Noted that for training COCO dataset, we use ImageNet pretrained weights to initialize both two backbones.

tomtomsa commented 1 year ago

@chenqi1126 I see, thank you for your reply.