ResNeSt: Split-Attention Networks

19 Apr 2020Hang ZhangChongruo WuZhongyue ZhangYi ZhuZhi ZhangHaibin LinYue SunTong HeJonas MuellerR. ManmathaMu LiAlexander Smola

While image classification models have recently continued to advance, most downstream applications such as object detection and semantic segmentation still employ ResNet variants as the backbone network due to their simple and modular structure. We present a simple and modular Split-Attention block that enables attention across feature-map groups... (read more)

PDF Abstract

Results from the Paper


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK USES EXTRA
TRAINING DATA
LEADERBOARD
Semantic Segmentation ADE20K ResNeSt-101 Validation mIoU 46.91 # 3
Semantic Segmentation ADE20K ResNeSt-269 Validation mIoU 47.60 # 2
Semantic Segmentation ADE20K ResNeSt-200 Validation mIoU 48.36 # 1
Semantic Segmentation ADE20K val ResNeSt-200 mIoU 48.36 # 1
Semantic Segmentation ADE20K val ResNeSt-269 mIoU 47.60 # 2
Semantic Segmentation ADE20K val ResNeSt-101 mIoU 46.91 # 4
Semantic Segmentation Cityscapes test ResNeSt200 Mean IoU (class) 83.3% # 8
Semantic Segmentation Cityscapes val ResNeSt200 mIoU 82.7% # 1
Object Detection COCO minival ResNeSt-200 (multi-scale) box AP 52.47 # 3
AP50 71.00 # 4
AP75 57.07 # 3
APS 36.80 # 3
APM 56.36 # 2
APL 66.29 # 3
Instance Segmentation COCO minival ResNeSt-200 (multi-scale) mask AP 46.25 # 1
Instance Segmentation COCO minival ResNeSt-101 (single-scale) mask AP 41.56 # 5
Instance Segmentation COCO minival ResNeSt-200 (single-scale) mask AP 44.21 # 3
Object Detection COCO minival ResNeSt-200 (single-scale) box AP 50.54 # 6
AP50 68.78 # 5
AP75 55.17 # 5
APM 54.2 # 5
APL 63.9 # 5
Panoptic Segmentation COCO panoptic ResNeSt-200 PQ 47.9 # 1
Instance Segmentation COCO test-dev ResNeSt-200 (multi-scale) mask AP 47.1 # 1
AP50 70.2 # 3
AP75 51.5 # 2
APS 30.0 # 2
APM 49.6 # 1
APL 60.6 # 2
Object Detection COCO test-dev ResNeSt-200DCN (multi-scale) box AP 53.3 # 4
AP50 72.0 # 4
AP75 58.0 # 4
APS 35.1 # 6
APM 56.2 # 4
APL 66.8 # 2
Image Classification ImageNet ResNeSt-101 Top 1 Accuracy 83.0% # 33
Image Classification ImageNet ResNeSt-269 Top 1 Accuracy 84.5% # 20
Image Classification ImageNet ResNeSt-200 Top 1 Accuracy 83.9% # 26
Semantic Segmentation PASCAL Context ResNeSt-101 mIoU 56.5 # 3
Semantic Segmentation PASCAL Context ResNeSt-200 mIoU 58.4 # 2
Semantic Segmentation PASCAL Context ResNeSt-269 mIoU 58.9 # 1

Methods used in the Paper