A ConvNet for the 2020s

Open paper
Benchmark Model Rank Results
classification-on-indlConvNext#1Average Recall: 93.47%
domain-generalization-on-imagenet-aConvNeXt-XL (Im21k, 384)#9Top-1 accuracy %: 69.3
domain-generalization-on-imagenet-cConvNeXt-XL (Im21k) (augmentation overlap with ImageNet-C)#12mean Corruption Error (mCE): 38.8Number of params: 350M
domain-generalization-on-imagenet-rConvNeXt-XL (Im21k, 384)#7Top-1 Error Rate: 31.8
domain-generalization-on-imagenet-sketchConvNeXt-XL (Im21k, 384)#3Top-1 accuracy: 55.0
domain-generalization-on-vizwizConvNeXt-B#2Accuracy - All Images: 53.5Accuracy - Corrupted Images: 46.9…
image-classification-on-imagenetAdlik-ViT-SG+Swin_large+Convnext_xlarge(384)#51Top 1 Accuracy: 88.36%Number of params: 1827M
image-classification-on-imagenetConvNeXt-XL (ImageNet-22k)#68Top 1 Accuracy: 87.8%Number of params: 350MGFLOPs: 179
image-classification-on-imagenetConvNeXt-L (384 res)#216Top 1 Accuracy: 85.5%Number of params: 198MGFLOPs: 101
image-classification-on-imagenetConvNeXt-T#557Top 1 Accuracy: 82.1%Number of params: 29MGFLOPs: 4.5
object-detection-on-coco-oConvNeXt-XL (Cascade Mask R-CNN)#7Average mAP: 37.5Effective Robustness: 12.68
semantic-segmentation-on-ade20kConvNeXt-XL++#67Validation mIoU: 54Params (M): 391GFLOPs (512 x 512): 3335
semantic-segmentation-on-ade20kConvNeXt-L++#72Validation mIoU: 53.7Params (M): 235GFLOPs (512 x 512): 2458
semantic-segmentation-on-ade20kConvNeXt-B++#80Validation mIoU: 53.1Params (M): 122GFLOPs (512 x 512): 1828
semantic-segmentation-on-ade20kConvNeXt-B#120Validation mIoU: 49.9Params (M): 122GFLOPs (512 x 512): 1170
semantic-segmentation-on-ade20kConvNeXt-S#127Validation mIoU: 49.6Params (M): 82GFLOPs (512 x 512): 1027
semantic-segmentation-on-ade20kConvNeXt-T#167Validation mIoU: 46.7Params (M): 60GFLOPs (512 x 512): 939
semantic-segmentation-on-imagenet-sConvNext-Tiny (P4, 224x224, SUP)#11mIoU (val): 48.7mIoU (test): 48.8

Rank counts only results with a code link.