MOAT: Alternating Mobile Convolution and Attention Brings Strong Vision Models

Open paper
Benchmark Model Rank Results
image-classification-on-imagenet-v2MOAT-4 (IN-22K pretraining)#10Top 1 Accuracy: 81.5
image-classification-on-imagenet-v2MOAT-3 (IN-22K pretraining)#14Top 1 Accuracy: 80.6
image-classification-on-imagenet-v2MOAT-2 (IN-22K pretraining)#18Top 1 Accuracy: 79.3
image-classification-on-imagenet-v2MOAT-1 (IN-22K pretraining)#20Top 1 Accuracy: 78.4
instance-segmentation-on-coco-minivalMOAT-3 (IN-22K pretraining, single-scale)#25mask AP: 50.3
instance-segmentation-on-coco-minivalMOAT-2 (IN-22K pretraining, single-scale)#30mask AP: 49.3
instance-segmentation-on-coco-minivalMOAT-1 (IN-1K pretraining, single-scale)#31mask AP: 49.0
instance-segmentation-on-coco-minivalMOAT-0 (IN-1K pretraining, single-scale)#46mask AP: 47.4
instance-segmentation-on-coco-minivaltiny-MOAT-3 (IN-1K pretraining, single-scale)#49mask AP: 47.0
instance-segmentation-on-coco-minivaltiny-MOAT-2 (IN-1K pretraining, single-scale)#61mask AP: 45.0
instance-segmentation-on-coco-minivaltiny-MOAT-1 (IN-1K pretraining, single-scale)#64mask AP: 44.6
instance-segmentation-on-coco-minivaltiny-MOAT-0 (IN-1K pretraining, single-scale)#76mask AP: 43.3
image-classification-on-imagenetMOAT-4 22K+1K#27Top 1 Accuracy: 89.1%Number of params: 483.2MGFLOPs: 648.5
image-classification-on-imagenetMOAT-3 1K only#137Top 1 Accuracy: 86.7%Number of params: 190MGFLOPs: 271
image-classification-on-imagenetMOAT-0 1K only#477Top 1 Accuracy: 83.3%Number of params: 27.8MGFLOPs: 5.7
semantic-segmentation-on-ade20kMOAT-4 (IN-22K pretraining, single-scale)#32Validation mIoU: 57.6Params (M): 496
semantic-segmentation-on-ade20kMOAT-3 (IN-22K pretraining, single-scale)#40Validation mIoU: 56.5Params (M): 198
semantic-segmentation-on-ade20kMOAT-2 (IN-22K pretraining, single-scale)#59Validation mIoU: 54.7Params (M): 81
semantic-segmentation-on-ade20ktiny-MOAT-3 (IN-1K pretraining, single scale)#180Validation mIoU: 47.5Params (M): 24
semantic-segmentation-on-ade20ktiny-MOAT-2 (IN-1K pretraining, single scale)#211Validation mIoU: 44.9Params (M): 13
semantic-segmentation-on-ade20ktiny-MOAT-1 (IN-1K pretraining, single scale)#226Validation mIoU: 43.1Params (M): 8
semantic-segmentation-on-ade20ktiny-MOAT-0 (IN-1K pretraining, single scale)#231Validation mIoU: 41.2Params (M): 6
object-detection-on-coco-val2017MOAT-3 (IN-22K pretraining, single-scale)#40box AP: 59.2
object-detection-on-coco-val2017MOAT-2 (IN-22K pretraining, single-scale)#49box AP: 58.5
object-detection-on-coco-val2017MOAT-1 (IN-1K pretraining, single-scale)#64box AP: 57.7
object-detection-on-coco-val2017MOAT-0 (IN-1K pretraining, single-scale)#92box AP: 55.9
object-detection-on-coco-val2017tiny-MOAT-3 (IN-1K pretraining, single-scale)#104box AP: 55.2
object-detection-on-coco-val2017tiny-MOAT-2 (IN-1K pretraining, single-scale)#154box AP: 53.0
object-detection-on-coco-val2017tiny-MOAT-1 (IN-1K pretraining, single-scale)#177box AP: 51.9
object-detection-on-coco-val2017tiny-MOAT-0 (IN-1K pretraining, single-scale)#207box AP: 50.5

Rank counts only results with a code link.