MixMAE: Mixed and Masked Autoencoder for Efficient Pretraining of Hierarchical Vision Transformers

Open paper
Benchmark Model Rank Results
image-classification-on-places365MixMIM-L(ViT-L)#2Top 1 Accuracy: 60.3
image-classification-on-places365MixMIM-B (ViT)#5Top 1 Accuracy: 58.9
image-classification-on-inaturalist-2019MixMIM-L#3Top-1 Accuracy: 83.9
image-classification-on-places205MixMIM-L#3Top 1 Accuracy: 69.3
image-classification-on-places205MixMIM-B#5Top 1 Accuracy: 68.3
image-classification-on-inaturalist-2018MixMIM-L#14Top-1 Accuracy: 80.3%
image-classification-on-inaturalist-2018MixMIM-B#20Top-1 Accuracy: 77.5%
semantic-segmentation-on-ade20k-valMixMIM-L#36mIoU: 53.8
semantic-segmentation-on-ade20k-valMixMIM-B#50mIoU: 50.3
object-detection-on-coco-val2017MixMAE (Swin-L, 600ep, Mask R-CNN)#124box AP: 54.3Params (M): 319
object-detection-on-coco-val2017MixMAE (Swin-B/W14, 600ep, Mask R-CNN)#163box AP: 52.7Params (M): 110
image-classification-on-imagenetMixMIM-B#276Top 1 Accuracy: 85.1%Number of params: 88MGFLOPs: 16.3

Rank counts only results with a code link.