人体检测器结果迭代

实验报告

前言

• 记录一下模型迭代的过程
• 涉及Yolo-FastestV1、Yolo-FastestV2、Yolov4-Tiny、Yolov5、YoloX等模型

模型选型

• 基础模型库表格（给组内看）

• 基础模型库表格（自己看）

Yolo-Fastest Human

• 对爆炸、火焰、烟画面误检多
• 人员密集时，对人群标注一个大框

class_id = 0, name = person, ap = 41.79%     (TP = 59817, FP = 62063)
 for conf_thresh = 0.25, precision = 0.49, recall = 0.45, F1-score = 0.47 
 for conf_thresh = 0.25, TP = 59817, FP = 62063, FN = 72779, average IoU = 36.59 % 
 IoU threshold = 50 %, used Area-Under-Curve for each unique Recall 
 mean average precision (mAP@0.50) = 0.417902, or 41.79 % 

yolov4-tiny-IR-head-416x416

 calculation mAP (mean average precision)...
 Detection layer: 30 - type = 27 
 Detection layer: 37 - type = 27 
388
 detections_count = 847, unique_truth_count = 2420  
class_id = 0, name = head, ap = 21.67%           (TP = 352, FP = 18) 
 for conf_thresh = 0.25, precision = 0.95, recall = 0.15, F1-score = 0.25 
 for conf_thresh = 0.25, TP = 352, FP = 18, FN = 2068, average IoU = 74.58 % 
 IoU threshold = 50 %, used Area-Under-Curve for each unique Recall 
 mean average precision (mAP@0.50) = 0.216670, or 21.67 % 
Total Detection Time: 5 Seconds

• ./yolov4_tiny_head_IR_cfg/backup_start/yolov4-tiny-IR-head-416x416_best.weights

 calculation mAP (mean average precision)...
 Detection layer: 30 - type = 27 
 Detection layer: 37 - type = 27 
388
 detections_count = 953, unique_truth_count = 2420  
class_id = 0, name = head, ap = 23.68%           (TP = 426, FP = 20) 
 for conf_thresh = 0.25, precision = 0.96, recall = 0.18, F1-score = 0.30 
 for conf_thresh = 0.25, TP = 426, FP = 20, FN = 1994, average IoU = 75.11 % 
 IoU threshold = 50 %, used Area-Under-Curve for each unique Recall 
 mean average precision (mAP@0.50) = 0.236756, or 23.68 % 
Total Detection Time: 4 Seconds

0924 RGB_human_head_car_v1.0

7368
 detections_count = 793939, unique_truth_count = 220056  
class_id = 0, name = person, ap = 25.00%         (TP = 36335, FP = 40491) 
class_id = 1, name = head, ap = 15.86%           (TP = 24146, FP = 61169) 
class_id = 2, name = car, ap = 16.51%            (TP = 1122, FP = 4569) 
 for conf_thresh = 0.25, precision = 0.37, recall = 0.28, F1-score = 0.32 
 for conf_thresh = 0.25, TP = 61603, FP = 106229, FN = 158453, average IoU = 26.33 % 
 IoU threshold = 50 %, used Area-Under-Curve for each unique Recall 
 mean average precision (mAP@0.50) = 0.191250, or 19.12 % 
Total Detection Time: 184 Seconds
Set -points flag:
 `-points 101` for MS COCO 
 `-points 11` for PascalVOC 2007 (uncomment `difficult` in voc.data) 
 `-points 0` (AUC) for ImageNet, PascalVOC 2010-2012, your custom dataset
 mean_average_precision (mAP@0.5) = 0.191250 
85908
 detections_count = 5484063, unique_truth_count = 1174038  
class_id = 0, name = person, ap = 31.15%         (TP = 234239, FP = 242316) 
class_id = 1, name = head, ap = 32.05%           (TP = 180216, FP = 240515) 
class_id = 2, name = car, ap = 29.04%            (TP = 30056, FP = 37160) 
 for conf_thresh = 0.25, precision = 0.46, recall = 0.38, F1-score = 0.42 
 for conf_thresh = 0.25, TP = 444511, FP = 519991, FN = 729527, average IoU = 34.25 % 
 IoU threshold = 50 %, used Area-Under-Curve for each unique Recall 
 mean average precision (mAP@0.50) = 0.307496, or 30.75 % 
Total Detection Time: 1007 Seconds
class_id = 0, name = person, ap = 27.14%         (TP = 38610, FP = 39395) 
class_id = 1, name = head, ap = 22.47%           (TP = 31045, FP = 56324) 
class_id = 2, name = car, ap = 23.11%            (TP = 2336, FP = 3504) 
 for conf_thresh = 0.25, precision = 0.42, recall = 0.31, F1-score = 0.36 
 for conf_thresh = 0.25, TP = 71991, FP = 99223, FN = 158505, average IoU = 30.88 % 
 IoU threshold = 50 %, used Area-Under-Curve for each unique Recall 
 mean average precision (mAP@0.50) = 0.242372, or 24.24 % 
Total Detection Time: 200 Seconds

0929 RGB_human_head_car_v2.0

 detections_count = 553895, unique_truth_count = 122198  
class_id = 0, name = person, ap = 24.69%         (TP = 37161, FP = 58262) 
class_id = 1, name = head, ap = 74.99%           (TP = 4337, FP = 3350) 
class_id = 2, name = car, ap = 28.02%            (TP = 1006, FP = 1583) 
 for conf_thresh = 0.25, precision = 0.40, recall = 0.35, F1-score = 0.37 
 for conf_thresh = 0.25, TP = 42504, FP = 63195, FN = 79694, average IoU = 30.70 % 
 IoU threshold = 50 %, used Area-Under-Curve for each unique Recall 
 mean average precision (mAP@0.50) = 0.425679, or 42.57 % 
Total Detection Time: 361 Seconds

0924 crowd human head

/ssd01/wanghuijiao/pose_detector02/crowdhuman_head_person.sh

4372
 detections_count = 385365, unique_truth_count = 206230  
class_id = 0, name = head, ap = 33.07%           (TP = 31763, FP = 4845) 
class_id = 1, name = person, ap = 53.16%         (TP = 52055, FP = 16293) 
 for conf_thresh = 0.25, precision = 0.80, recall = 0.41, F1-score = 0.54 
 for conf_thresh = 0.25, TP = 83818, FP = 21138, FN = 122412, average IoU = 60.57 % 
 IoU threshold = 50 %, used Area-Under-Curve for each unique Recall 
 mean average precision (mAP@0.50) = 0.431143, or 43.11 % 
Total Detection Time: 173 Seconds

0926 APs

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.111
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.253
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.088
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.022
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.129
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.258
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.045
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.153
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.182
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.040
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.209
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.371