YOLOV3/ YOLOV4/Centernet 指标比较

COCO数据集合，1080Ti：

 

yolov3		yolo4		centernet
fps	ap	fps	ap	fps	ap
20	33	43	43	28	41

 

yolov4：自己数据集合测试结果：5000 valid数据，其中包括BDD80%，自采数据20%。

./darknet detector map data/whl_yolov4.data cfg/cfgs_whl/yolov4-custom.cfg backup/yolov4-custom_last.weight

class_id = 0, name = car, ap = 70.56%        (TP = 31806, FP = 12516)
class_id = 1, name = bus, ap = 55.24%        (TP = 495, FP = 343)
class_id = 2, name = person, ap = 54.52%        (TP = 3196, FP = 1246)
class_id = 3, name = bike, ap = 36.33%        (TP = 152, FP = 146)
class_id = 4, name = truck, ap = 54.24%        (TP = 858, FP = 583)
class_id = 5, name = motor, ap = 30.69%        (TP = 54, FP = 76)
class_id = 6, name = train, ap = 0.00%        (TP = 0, FP = 0)
class_id = 7, name = rider, ap = 40.53%        (TP = 148, FP = 145)
class_id = 8, name = sign, ap = 61.68%        (TP = 10482, FP = 4970)
class_id = 9, name = Red, ap = 73.51%        (TP = 459, FP = 251)
class_id = 10, name = Green, ap = 78.14%        (TP = 525, FP = 169)
class_id = 11, name = Yellow, ap = 60.27%        (TP = 17, FP = 1)
class_id = 12, name = Off, ap = 33.26%        (TP = 35, FP = 37)
class_id = 13, name = GreenHorizontal, ap = 0.00%        (TP = 0, FP = 0)
class_id = 14, name = RedHorizontal, ap = 0.00%        (TP = 0, FP = 0)
class_id = 15, name = YellowHorizontal, ap = 0.00%        (TP = 0, FP = 0)
class_id = 16, name = GreenLeft, ap = 50.39%        (TP = 16, FP = 13)
class_id = 17, name = RedLeft, ap = 64.29%        (TP = 106, FP = 80)
class_id = 18, name = YellowLeft, ap = 0.00%        (TP = 0, FP = 0)
class_id = 19, name = GreenRight, ap = 0.00%        (TP = 0, FP = 0)
class_id = 20, name = RedRight, ap = 0.00%        (TP = 0, FP = 0)
class_id = 21, name = YellowRight, ap = 0.00%        (TP = 0, FP = 0)
class_id = 22, name = GreenTurn, ap = 67.51%        (TP = 14, FP = 6)
class_id = 23, name = RedTurn, ap = 41.25%        (TP = 19, FP = 18)
class_id = 24, name = YellowTurn, ap = 0.00%        (TP = 0, FP = 0)
class_id = 25, name = Cone, ap = 65.88%        (TP = 142, FP = 62)
class_id = 26, name = TriangleWarningSign, ap = 21.08%        (TP = 0, FP = 1)

 for conf_thresh = 0.25, precision = 0.70, recall = 0.65, F1-score = 0.68
 for conf_thresh = 0.25, TP = 48524, FP = 20663, FN = 26059, average IoU = 54.38 %

 IoU threshold = 50 %, used Area-Under-Curve for each unique Recall
 mean average precision (mAP@0.50) = 0.355323, or 35.53 %