Performance of Swin-L

[dawn-ech]

Hi @TempleX98, thanks for your awesome work.
Co-DINO with ViT-L achieves 66.0%AP. I would like to know the performance of Co-DINO with Swin-L. Thank you.

[TempleX98]

We don't have the result of O365 pretrained Co-DETR+Swin-L

[HITerStudy]

We don't have the result of O365 pretrained Co-DETR+Swin-L

The training pipeline for the 66.0% AP of Co-Dino(ViT-L) is same as that in this respository? such as image resolution and multi-scale range in data augmentation? Please offer some details, thanks!

[TempleX98]

We don't have the result of O365 pretrained Co-DETR+Swin-L

The training pipeline for the 66.0% AP of Co-Dino(ViT-L) is same as that in this respository? such as image resolution and multi-scale range in data augmentation? Please offer some details, thanks!

It's the same model as the released part. The image resolution for pretraining is [480, 1024]x1280. We will update the results and model details in the arxiv version soon.

[HITerStudy]

We don't have the result of O365 pretrained Co-DETR+Swin-L

The training pipeline for the 66.0% AP of Co-Dino(ViT-L) is same as that in this respository? such as image resolution and multi-scale range in data augmentation? Please offer some details, thanks!

It's the same model as the released part. The image resolution for pretraining is [480, 1024]x1280. We will update the results and model details in the arxiv version soon.

the pretraining dataset is obj365, it is the v1 version or obj365v2?

[TempleX98]

We don't have the result of O365 pretrained Co-DETR+Swin-L

The training pipeline for the 66.0% AP of Co-Dino(ViT-L) is same as that in this respository? such as image resolution and multi-scale range in data augmentation? Please offer some details, thanks!

It's the same model as the released part. The image resolution for pretraining is [480, 1024]x1280. We will update the results and model details in the arxiv version soon.

the pretraining dataset is obj365, it is the v1 version or obj365v2?

v2

[songyuc]

Hi, @TempleX98, have the model config and weight of 66.0 boxAP been released?

[TempleX98]

@songyuc Sorry, we have no plan to release the ViT-L model now. But we may release another O365 pretrained model in a few weeks.

[songyuc]

Thanks sincerely.

[TempleX98]

@dawn-ech @songyuc We have provided the O365 pre-trained Swin-L model

[ganliqiang]

hi ,i cannot find the pretrain model swin_large_patch4_window12_384_22k.pth so,the leatest model co_dino_5scale_swin_large_16e_o365tococo.pth is same as the pretrained model in the config file.is it right ?how i download the swin_large_patch4_window12_384_22k.pth.thanks a lot in advanced.

[TempleX98]

@ganliqiang https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_large_patch4_window12_384_22k.pth