Mask R-CNN

ICCV 2017 Kaiming HeGeorgia GkioxariPiotr DollárRoss Girshick

We present a conceptually simple, flexible, and general framework for object instance segmentation. Our approach efficiently detects objects in an image while simultaneously generating a high-quality segmentation mask for each instance... (read more)

PDF Abstract ICCV 2017 PDF ICCV 2017 Abstract

Code


Results from the Paper


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK RESULT BENCHMARK
Nuclear Segmentation Cell17 Mask R-CNN F1-score 0.8004 # 3
Dice 0.707 # 3
Hausdorff 12.6723 # 3
Panoptic Segmentation Cityscapes val Mask R-CNN+COCO PQth 54.0 # 12
Keypoint Detection COCO Mask R-CNN Validation AP 69.2 # 6
Test AP 63.1 # 8
Real-Time Object Detection COCO Mask R-CNN X-152-32x8d MAP 40.3 # 9
FPS 3 # 16
inference time (ms) 333 # 14
Real-Time Object Detection COCO minival Mask R-CNN X-101-FPN MAP 37.6 # 1
Object Detection COCO minival Mask R-CNN (ResNet-101-FPN) box AP 40.0 # 41
Object Detection COCO minival Mask R-CNN (ResNet-50-FPN) box AP 37.7 # 55
Object Detection COCO minival Mask R-CNN (ResNeXt-101-FPN) box AP 36.7 # 57
AP50 59.5 # 30
AP75 38.9 # 42
Real-Time Object Detection COCO minival Mask R-CNN X-152-32x8d APbb75 45.2 # 1
Keypoint Detection COCO test-challenge Mask R-CNN* AR 75.4 # 4
ARM 70.2 # 4
AP 68.9 # 5
AP50 89.2 # 4
AP75 75.2 # 4
APL 82.6 # 4
AR50 93.2 # 4
AR75 81.2 # 4
ARL 76.8 # 4
Instance Segmentation COCO test-dev Mask R-CNN (ResNeXt-101-FPN) mask AP 37.1 # 20
AP50 60.0 # 12
AP75 39.4 # 11
APS 16.9 # 14
APM 39.9 # 12
APL 53.5 # 9
Pose Estimation COCO test-dev Mask-RCNN AP 63.1 # 10
AP50 87.3 # 7
AP75 68.7 # 10
APL 71.4 # 8
Object Detection COCO test-dev Mask R-CNN (ResNet-101-FPN) box AP 38.2 # 69
AP50 60.3 # 57
AP75 41.7 # 71
APS 20.1 # 74
APM 41.1 # 67
APL 50.2 # 68
Object Detection COCO test-dev Mask R-CNN (ResNeXt-101-FPN) box AP 39.8 # 61
AP50 62.3 # 48
AP75 43.4 # 64
APS 22.1 # 64
APM 43.2 # 58
APL 51.2 # 63
Keypoint Detection COCO test-dev Mask R-CNN APL 71.4 # 10
APM 57.8 # 12
AP50 87.3 # 9
AP75 68.7 # 11
Multi-Person Pose Estimation CrowdPose Mask R-CNN mAP @0.5:0.95 60.3 # 4
Multi-tissue Nucleus Segmentation Kumar Mask R-CNN (e) Dice 0.760 # 13
Hausdorff Distance (mm) 50.9 # 11
Multi-Human Parsing MHP v1.0 Mask R-CNN AP 0.5 52.68% # 2
Multi-Human Parsing MHP v2.0 Mask R-CNN AP 0.5 14.9% # 3

Methods used in the Paper