yolov4-tiny训练时数据均衡与不均衡的对比实验

东方耀

yolov4-tiny训练时数据均衡与不均衡的对比实验
训练(先不用3层的 用传统的yolov4-tiny  608*608)：
batch=64
subdivisions=16  max_batches = 10000
./darknet detector train backup_person_vehicle/person_vehicle_mini.data backup_person_vehicle/yolov4-tiny_person_vehicle.cfg backup/yolov4-tiny.conv.29 -map
统一用2951张val set看指标：
./darknet detector map backup_person_vehicle/person_vehicle.data backup_person_vehicle/yolov4-tiny_person_vehicle.cfg backup_person_vehicle_v4/yolov4-tiny_person_vehicle_best.weights
class_id = 0, name = person, ap = 59.18%            (TP = 6116, FP = 3613)
class_id = 1, name = car, ap = 47.50%            (TP = 940, FP = 833)
class_id = 2, name = motorbike, ap = 40.96%            (TP = 119, FP = 80)
class_id = 3, name = bus, ap = 55.21%            (TP = 132, FP = 83)
class_id = 4, name = truck, ap = 32.05%            (TP = 135, FP = 150)
(mAP@0.50) = 0.46


看看数据平衡的情况下{只改了训练数据 val集没有变}：
./darknet detector train backup_person_vehicle/person_vehicle_mini_balance.data backup_person_vehicle/yolov4-tiny_person_vehicle.cfg backup/yolov4-tiny.conv.29 -map
统一用2951张val set看指标：
./darknet detector map backup_person_vehicle/person_vehicle.data backup_person_vehicle/yolov4-tiny_person_vehicle.cfg backup_person_vehicle_v4_balance/yolov4-tiny_person_vehicle_balance_best.weights
class_id = 0, name = person, ap = 55.60%            (TP = 5598, FP = 3591)
class_id = 1, name = car, ap = 51.96%            (TP = 1017, FP = 913)
class_id = 2, name = motorbike, ap = 46.16%            (TP = 155, FP = 124)
class_id = 3, name = bus, ap = 65.12%            (TP = 178, FP = 113)
class_id = 4, name = truck, ap = 36.83%            (TP = 184, FP = 279)
(mAP@0.50) = 0.511
结论：训练的数据经过平衡化处理后【对比实验】，map提高了5个点

leironh

666666666666666666666666

leironh

6666666666666666

leironh

笔记写的挺好的，点个赞

leironh

目标检测之voc格式的数据集转换为csv格式

leironh

目标检测之voc格式的数据集转换为csv格式目标检测之voc格式的数据集转换为csv格式目标检测之voc格式的数据集转换为csv格式目标检测之voc格式的数据集转换为csv格式

leironh

目标检测之voc格式的数据集转换为csv格式目标检测之voc格式的数据集转换为csv格式

leironh

目标检测之coco与voc格式的数据相互转换，并验证coco格式的脚本目标检测之coco与voc格式的数据相互转换，并验证coco格式的脚本目标检测之coco与voc格式的数据相互转换，并验证coco格式的脚本