A Ranking-based, Balanced Loss Function Unifying Classification and Localisation in Object Detection

Benchmark Model Rank Results
object-detection-on-cocoaLRP Loss (ResNext-101-64x4d, DCN, multiscale test)#89box mAP: 50.2AP50: 70.3AP75: 53.9APS: 32.0APM: 53.1APL: 63.0
object-detection-on-cocoaLRP Loss (ResNext-101-64x4d, DCN, single scale)#99box mAP: 48.9AP50: 69.3AP75: 52.5APS: 30.8APM: 51.5APL: 62.1
object-detection-on-cocoaLRP Loss (ResNext-101-64x4d, single scale)#113box mAP: 47.8AP50: 68.4AP75: 51.1APS: 30.2APM: 50.8APL: 59.1
object-detection-on-cocoaLRP Loss (ResNext-101, DCN, 500 scale)#142box mAP: 44.6AP50: 65.0AP75: 47.5APS: 24.6APM: 48.1APL: 58.3
object-detection-on-coco-minivalFaster R-CNN+aLRP Loss (ResNet-50, 500 scale)#175box AP: 40.7AP50: 60.7AP75: 43.3
object-detection-on-coco-minivalRetinaNet+aLRP Loss (ResNet-50, 500 scale)#182box AP: 40.2AP50: 60.3AP75: 42.3
object-detection-on-coco-minivalFoveaBox+aLRP Loss (ResNet-50, 500 scale)#185box AP: 39.7AP50: 58.8AP75: 41.5