Why AP values decrease after k-means++?

Viewed 10

I used k-means++ to size the anchor boxes to give more AP. As a result, the AP is lower than before applying k-means++. Please let me know if there is any cause for this. Thank you in advance.

before: anchor box (10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326)

detections_count = 13032, unique_truth_count = 2164                                                                            
class_id = 0, name = Can, ap = 90.15%            (TP = 240, FP = 25)                                                                 
class_id = 1, name = Bottle, ap = 96.47%         (TP = 898, FP = 60)                                                                  
class_id = 2, name = Box, ap = 74.26%            (TP = 50, FP = 21)                                                                  
class_id = 3, name = Cigarette, ap = 59.19%      (TP = 506, FP = 323)                                                                                                                                                                                                        for conf_thresh = 0.25, precision = 0.80, recall = 0.78, F1-score = 0.79                                                              
for conf_thresh = 0.25, TP = 1694, FP = 429, FN = 470, average IoU = 66.07 %                                                                                                                                                                                                IoU threshold = 50 %, used Area-Under-Curve for each unique Recall                                                                    
mean average precision (mAP@0.50) = 0.800176, or 80.02 %    

after:anchor box (42, 68, 114, 80, 72,159, 187,149, 128,312, 277,330)

 detections_count = 3695, unique_truth_count = 2164                                                                                  
 class_id = 0, name = Can, ap = 91.05%            (TP = 246, FP = 11)                                                                 
 class_id = 1, name = Bottle, ap = 93.78%         (TP = 862, FP = 46)                                                                 
 class_id = 2, name = Box, ap = 76.94%            (TP = 42, FP = 11)                                                                   
 class_id = 3, name = Cigarette, ap = 37.37%      (TP = 184, FP = 57)                                                                                                                                                                                                         for conf_thresh = 0.25, precision = 0.91, recall = 0.62, F1-score = 0.74                                                              
for conf_thresh = 0.25, TP = 1334, FP = 125, FN = 830, average IoU = 78.62 %                                                                                                                                                                                                IoU threshold = 50 %, used Area-Under-Curve for each unique Recall                                                                    
mean average precision (mAP@0.50) = 0.747871, or 74.79 % 
0 Answers
Related