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

We propose average Localisation-Recall-Precision (aLRP), a unified, bounded, balanced and ranking-based loss function for both classification and localisation tasks in object detection. aLRP extends the Localisation-Recall-Precision (LRP) performance metric (Oksuz et al., 2018) inspired from how Average Precision (AP) Loss extends precision to a ranking-based loss function for classification (Chen et al., 2020)... aLRP has the following distinct advantages: (i) aLRP is the first ranking-based loss function for both classification and localisation tasks. (ii) Thanks to using ranking for both tasks, aLRP naturally enforces high-quality localisation for high-precision classification. (iii) aLRP provides provable balance between positives and negatives. (iv) Compared to on average $\sim$6 hyperparameters in the loss functions of state-of-the-art detectors, aLRP Loss has only one hyperparameter, which we did not tune in practice. On the COCO dataset, aLRP Loss improves its ranking-based predecessor, AP Loss, up to around $5$ AP points, achieves $48.9$ AP without test time augmentation and outperforms all one-stage detectors. Code available at: https://github.com/kemaloksuz/aLRPLoss . read more

PDF Abstract NeurIPS 2020 PDF NeurIPS 2020 Abstract

Datasets


Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Object Detection COCO minival FoveaBox+aLRP Loss (ResNet-50, 500 scale) box AP 39.7 # 112
AP50 58.8 # 77
AP75 41.5 # 72
Object Detection COCO minival RetinaNet+aLRP Loss (ResNet-50, 500 scale) box AP 40.2 # 109
AP50 60.3 # 62
AP75 42.3 # 68
Object Detection COCO minival Faster R-CNN+aLRP Loss (ResNet-50, 500 scale) box AP 40.7 # 102
AP50 60.7 # 61
AP75 43.3 # 65
Object Detection COCO test-dev aLRP Loss (ResNext-101-64x4d, single scale) box AP 47.8 # 68
AP50 68.4 # 43
AP75 51.1 # 56
APS 30.2 # 45
APM 50.8 # 50
APL 59.1 # 59
Hardware Burden None # 1
Operations per network pass None # 1
Object Detection COCO test-dev aLRP Loss (ResNext-101-64x4d, DCN, single scale) box AP 48.9 # 59
AP50 69.3 # 38
AP75 52.5 # 49
APS 30.8 # 42
APM 51.5 # 48
APL 62.1 # 36
Hardware Burden None # 1
Operations per network pass None # 1
Object Detection COCO test-dev aLRP Loss (ResNext-101, DCN, 500 scale) box AP 44.6 # 97
AP50 65.0 # 69
AP75 47.5 # 87
APS 24.6 # 95
APM 48.1 # 72
APL 58.3 # 62
Hardware Burden None # 1
Operations per network pass None # 1
Object Detection COCO test-dev aLRP Loss (ResNext-101-64x4d, DCN, multiscale test) box AP 50.2 # 51
AP50 70.3 # 29
AP75 53.9 # 39
APS 32.0 # 33
APM 53.1 # 35
APL 63.0 # 32
Hardware Burden None # 1
Operations per network pass None # 1

Methods


No methods listed for this paper. Add relevant methods here