Papers › MOAT: Alternating Mobile Convolution and Attention Brings Strong Vision Models

MOAT: Alternating Mobile Convolution and Attention Brings Strong Vision Models

4 Oct 2022arXiv:2210.01820archive 2025-07-28

Chenglin Yang, Siyuan Qiao, Qihang Yu, Xiaoding Yuan, Yukun Zhu, Alan Yuille, Hartwig Adam, Liang-Chieh Chen

This paper presents MOAT, a family of neural networks that build on top of MObile convolution (i.e., inverted residual blocks) and ATtention. Unlike the current works that stack separate mobile convolution and transformer blocks, we effectively merge them into a MOAT block. Starting with a standard Transformer block, we replace its multi-layer perceptron with a mobile convolution block, and further reorder it before the self-attention operation. The mobile convolution block not only enhances the network representation capacity, but also produces better downsampled features. Our conceptually simple MOAT networks are surprisingly effective, achieving 89.1% / 81.5% top-1 accuracy on ImageNet-1K / ImageNet-1K-V2 with ImageNet22K pretraining. Additionally, MOAT can be seamlessly applied to downstream tasks that require large resolution inputs by simply converting the global attention to window attention. Thanks to the mobile convolution that effectively exchanges local information between pixels (and thus cross-windows), MOAT does not need the extra window-shifting mechanism. As a result, on COCO object detection, MOAT achieves 59.2% box AP with 227M model parameters (single-scale inference, and hard NMS), and on ADE20K semantic segmentation, MOAT attains 57.6% mIoU with 496M model parameters (single-scale inference). Finally, the tiny-MOAT family, obtained by simply reducing the channel sizes, also surprisingly outperforms several mobile-specific transformer-based models on ImageNet. The tiny-MOAT family is also benchmarked on downstream tasks, serving as a baseline for the community. We hope our simple yet effective MOAT will inspire more seamless integration of convolution and self-attention. Code is publicly available.

PaperPDFCode

In Syntology Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

Code

google-research/deeplab2 officialmentioned in papertfApache-2.0 report
RooKichenn/pytorch-MOAT mentioned on GitHubpytorch report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

Image ClassificationInstance SegmentationObject DetectionSemantic Segmentation

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
Image Classification ImageNet MOAT-4 22K+1K GFLOPs 648.5 #25 of 1060 Archive leaderboard report
Image Classification ImageNet MOAT-4 22K+1K Number of params 483.2M #25 of 1060 Archive leaderboard report
Image Classification ImageNet MOAT-4 22K+1K Top 1 Accuracy 89.1% #25 of 1060 Archive leaderboard report
Image Classification ImageNet MOAT-3 1K only GFLOPs 271 #129 of 1060 Archive leaderboard report
Image Classification ImageNet MOAT-3 1K only Number of params 190M #129 of 1060 Archive leaderboard report
Image Classification ImageNet MOAT-3 1K only Top 1 Accuracy 86.7% #129 of 1060 Archive leaderboard report
Image Classification ImageNet MOAT-0 1K only GFLOPs 5.7 #439 of 1060 Archive leaderboard report
Image Classification ImageNet MOAT-0 1K only Number of params 27.8M #439 of 1060 Archive leaderboard report
Image Classification ImageNet MOAT-0 1K only Top 1 Accuracy 83.3% #439 of 1060 Archive leaderboard report
Image Classification ImageNet V2 MOAT-4 (IN-22K pretraining) Top 1 Accuracy 81.5 #8 of 33 Archive leaderboard report
Image Classification ImageNet V2 MOAT-3 (IN-22K pretraining) Top 1 Accuracy 80.6 #10 of 33 Archive leaderboard report
Image Classification ImageNet V2 MOAT-2 (IN-22K pretraining) Top 1 Accuracy 79.3 #11 of 33 Archive leaderboard report
Image Classification ImageNet V2 MOAT-1 (IN-22K pretraining) Top 1 Accuracy 78.4 #12 of 33 Archive leaderboard report
Instance Segmentation COCO minival MOAT-3 (IN-22K pretraining, single-scale) mask AP 50.3 #25 of 93 Archive leaderboard report
Instance Segmentation COCO minival MOAT-2 (IN-22K pretraining, single-scale) mask AP 49.3 #28 of 93 Archive leaderboard report
Instance Segmentation COCO minival MOAT-1 (IN-1K pretraining, single-scale) mask AP 49.0 #29 of 93 Archive leaderboard report
Instance Segmentation COCO minival MOAT-0 (IN-1K pretraining, single-scale) mask AP 47.4 #38 of 93 Archive leaderboard report
Instance Segmentation COCO minival tiny-MOAT-3 (IN-1K pretraining, single-scale) mask AP 47.0 #41 of 93 Archive leaderboard report
Instance Segmentation COCO minival tiny-MOAT-2 (IN-1K pretraining, single-scale) mask AP 45.0 #49 of 93 Archive leaderboard report
Instance Segmentation COCO minival tiny-MOAT-1 (IN-1K pretraining, single-scale) mask AP 44.6 #51 of 93 Archive leaderboard report
Instance Segmentation COCO minival tiny-MOAT-0 (IN-1K pretraining, single-scale) mask AP 43.3 #61 of 93 Archive leaderboard report
Object Detection COCO (Common Objects in Context) MOAT-3 22K+1K box AP 59.2 #1 of 3 Archive leaderboard report
Object Detection COCO (Common Objects in Context) MOAT-2 box AP 58.5 #2 of 3 Archive leaderboard report
Object Detection COCO minival MOAT-3 (IN-22K pretraining, single-scale) box AP 59.2 #31 of 220 Archive leaderboard report
Object Detection COCO minival MOAT-2 (IN-22K pretraining, single-scale) box AP 58.5 #35 of 220 Archive leaderboard report
Object Detection COCO minival MOAT-1 (IN-1K pretraining, single-scale) box AP 57.7 #38 of 220 Archive leaderboard report
Object Detection COCO minival MOAT-0 (IN-1K pretraining, single-scale) box AP 55.9 #47 of 220 Archive leaderboard report
Object Detection COCO minival tiny-MOAT-3 (IN-1K pretraining, single-scale) box AP 55.2 #50 of 220 Archive leaderboard report
Object Detection COCO minival tiny-MOAT-2 (IN-1K pretraining, single-scale) box AP 53.0 #63 of 220 Archive leaderboard report
Object Detection COCO minival tiny-MOAT-1 (IN-1K pretraining, single-scale) box AP 51.9 #69 of 220 Archive leaderboard report
Object Detection COCO minival tiny-MOAT-0 (IN-1K pretraining, single-scale) box AP 50.5 #78 of 220 Archive leaderboard report
Semantic Segmentation ADE20K MOAT-4 (IN-22K pretraining, single-scale) Params (M) 496 #30 of 235 Archive leaderboard report
Semantic Segmentation ADE20K MOAT-4 (IN-22K pretraining, single-scale) Validation mIoU 57.6 #30 of 235 Archive leaderboard report
Semantic Segmentation ADE20K MOAT-3 (IN-22K pretraining, single-scale) Params (M) 198 #39 of 235 Archive leaderboard report
Semantic Segmentation ADE20K MOAT-3 (IN-22K pretraining, single-scale) Validation mIoU 56.5 #39 of 235 Archive leaderboard report
Semantic Segmentation ADE20K MOAT-2 (IN-22K pretraining, single-scale) Params (M) 81 #56 of 235 Archive leaderboard report
Semantic Segmentation ADE20K MOAT-2 (IN-22K pretraining, single-scale) Validation mIoU 54.7 #56 of 235 Archive leaderboard report
Semantic Segmentation ADE20K tiny-MOAT-3 (IN-1K pretraining, single scale) Params (M) 24 #163 of 235 Archive leaderboard report
Semantic Segmentation ADE20K tiny-MOAT-3 (IN-1K pretraining, single scale) Validation mIoU 47.5 #163 of 235 Archive leaderboard report
Semantic Segmentation ADE20K tiny-MOAT-2 (IN-1K pretraining, single scale) Params (M) 13 #198 of 235 Archive leaderboard report
Semantic Segmentation ADE20K tiny-MOAT-2 (IN-1K pretraining, single scale) Validation mIoU 44.9 #198 of 235 Archive leaderboard report
Semantic Segmentation ADE20K tiny-MOAT-1 (IN-1K pretraining, single scale) Params (M) 8 #212 of 235 Archive leaderboard report
Semantic Segmentation ADE20K tiny-MOAT-1 (IN-1K pretraining, single scale) Validation mIoU 43.1 #212 of 235 Archive leaderboard report
Semantic Segmentation ADE20K tiny-MOAT-0 (IN-1K pretraining, single scale) Params (M) 6 #216 of 235 Archive leaderboard report
Semantic Segmentation ADE20K tiny-MOAT-0 (IN-1K pretraining, single scale) Validation mIoU 41.2 #216 of 235 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

Absolute Position EncodingsAdamAttentionBPEConvolutionDense ConnectionsDropoutLabel SmoothingLayer NormalizationLinear LayerMulti-Head AttentionPosition-Wise Feed-Forward LayerResidual ConnectionSoftmaxTransformer

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections