Comprehensive results table

July 11, 2019 ยท View on GitHub

Results are sorted by their absolute performance under corruption.

ResNet-50 Backbone Track

In this section only models with a ResNet 50 Backbone with Feature Pyramid Networks are listed.

Coco

Object detection:

ModelBackbonebox AP cleanbox AP corr.box %
Cascade Mask R-CNNR-50-FPN41.220.750.2
Faster R-CNN CombinedR-50-FPN34.620.458.9
Cascade R-CNNR-50-FPN40.420.149.7
Mask R-CNNR-50-FPN37.318.750.1
Faster R-CNNR-50-FPN36.318.250.2
RetinaNetR-50-FPN35.617.850.1
Faster R-CNN StylizedR-50-FPN21.514.165.6

Instance Segmentation:

ModelBackbonemask AP cleanmask AP corr.mask %
Mask R-CNN CombinedR-50-FPN32.919.057.7
Cascade Mask R-CNNR-50-FPN35.717.649.3
Mask R-CNNR-50-FPN34.216.849.1
Mask R-CNN StylizesR-50-FPN30.513.264.1

Pascal VOC

Object detection:

ModelBackbonebox AP50 cleanbox AP50 corr.box %
Faster R-CNN CombinedR-50-FPN80.456.269.9
Faster R-CNN StylizedR-50-FPN68.050.073.5
Faster R-CNNR-50-FPN80.548.660.4

Cityscapes

Object detection:

ModelBackbonebox AP cleanbox AP corr.box %
Faster R-CNN CombinedR-50-FPN36.317.247.4
Faster R-CNN StylizedR-50-FPN28.514.751.5
Faster R-CNNR-50-FPN36.412.233.4
Mask R-CNNR-50-FPN37.511.731.1

Instance Segmentation:

ModelBackbonemask AP cleanmask AP corr.mask %
Mask R-CNN CombinedR-50-FPN32.114.946.3
Mask R-CNN StylizesR-50-FPN23.011.349.2
Mask R-CNNR-50-FPN32.710.030.5

Unrestricted Track

Any mdel independent of it's backbone can participate in this track.

Coco

Object detection:

ModelBackbonebox AP cleanbox AP corr.box %
Hybrid Task CascadeX-101-64x4d-FPN-DCN50.632.764.7
Faster R-CNNX-101-32x4d-FPN-DCN43.426.761.6
Faster R-CNNX-101-64x4d-FPN41.323.456.6
Mask R-CNNR-50-FPN-DCN41.123.356.7
Faster R-CNNR-50-FPN-DCN40.022.456.1
Faster R-CNNX-101-32x4d-FPN40.122.355.5
Faster R-CNNR-101-FPN38.520.954.2
Cascade Mask R-CNNR-50-FPN41.220.750.2
Faster R-CNN CombinedR-50-FPN34.620.458.9
Cascade R-CNNR-50-FPN40.420.149.7
Mask R-CNNR-50-FPN37.318.750.1
Faster R-CNNR-50-FPN36.318.250.2
RetinaNetR-50-FPN35.617.850.1
Faster R-CNN StylizedR-50-FPN21.514.165.6

Instance Segmentation:

ModelBackbonemask AP cleanmask AP corr.mask %
Hybrid Task CascadeX-101-64x4d-FPN-DCN43.828.164.0
Mask R-CNNR-50-FPN-DCN37.220.755.7
Mask R-CNN CombinedR-50-FPN32.919.057.7
Cascade Mask R-CNNR-50-FPN35.717.649.3
Mask R-CNNR-50-FPN34.216.849.1
Mask R-CNN StylizesR-50-FPN30.513.264.1

Pascal VOC

Object detection:

ModelBackbonebox AP50 cleanbox AP50 corr.box %
Faster R-CNN CombinedR-50-FPN80.456.269.9
Faster R-CNN StylizedR-50-FPN68.050.073.5
Faster R-CNNR-50-FPN80.548.660.4

Cityscapes

Object detection:

ModelBackbonebox AP cleanbox AP corr.box %
Faster R-CNN CombinedR-50-FPN36.317.247.4
Faster R-CNN StylizedR-50-FPN28.514.751.5
Faster R-CNNR-50-FPN36.412.233.4
Mask R-CNNR-50-FPN37.511.731.1

Instance Segmentation:

ModelBackbonemask AP cleanmask AP corr.mask %
Mask R-CNN CombinedR-50-FPN32.114.946.3
Mask R-CNN StylizesR-50-FPN23.011.349.2
Mask R-CNNR-50-FPN32.710.030.5

Methods