Deformable ConvNets v2: More Deformable, Better Results

Benchmark Model Rank Results
object-detection-on-coco-test-devDCNv2 (ResNet-101, multi-scale)–box mAP: 46.0AP50: 67.9AP75: 50.8APS: 27.8APM: 49.1APL: 59.5
object-detection-on-coco-val2017Faster R-CNN (ResNet-101, DCNv2)–box AP: 41.7APS: 22.2APM: 45.8APL: 58.7
object-detection-on-coco-val2017Mask R-CNN (ResNet-101, DCNv2)–box AP: 43.1