Rethinking the Optimization of Average Precision: Only Penalizing Negative Instances before Positive Ones is Enough

Benchmark Model Rank Results
image-retrieval-on-inaturalistPNP Loss#10R@1: 66.6R@16: 89.7R@32: 92.6
image-retrieval-on-sopPNP Loss#6R@1: 81.1
vehicle-re-identification-on-vehicleid-largePNP Loss#3Rank-1: 93.2Rank-5: 96.6
vehicle-re-identification-on-vehicleid-mediumPNP Loss#3Rank-1: 94.2Rank-5: 96.9
vehicle-re-identification-on-vehicleid-smallPNP Loss#3Rank-1: 95.5Rank-5: 97.8