70+epoch，验证集上P0.9，但是R只有0.03 about smart_construction HOT 12 CLOSED

sorry,I solved the problem,mainly for anchors and data labels,get the same score with author,Thank you.

from smart_construction.

same question,P is OK,but recall is low ,map0.5≈0.4

from smart_construction.

将 test.py 入参的默认值 conf_thres=0.4,改大一点，我测试的时候将其设置为 0.6

from smart_construction.

将 test.py 入参的默认值 conf_thres=0.4,改大一点，我测试的时候将其设置为 0.6

将你提供的helment_yolov5s.pt权重文件拿来测试，conf=0.6情况下，R还是只有0.4，还有其他地方需要修改吗

from smart_construction.

将 test.py 入参的默认值 conf_thres=0.4,改大一点，我测试的时候将其设置为 0.6

阈值越大 召回就更低了呀

from smart_construction.

labels.npg里类别显示均衡吗，这边显示很不均衡

from smart_construction.

@PeterH0323
whether coordinate calculation error to bbox of head, helmet?

from smart_construction.

@PeterH0323
146th epoch:
people bbox P and R were better，while P and R of head or helmet were worse.

as following:
1.training:
145/299 9.17G 0.04331 0.1072 0.004936 0.1554 7980 640: 99%|███████████████████████▋| 85/86 145/299 9.17G 0.04325 0.1071 0.004934 0.1552 1441 640: 99%|███████████████████████▋| 85/86 145/299 9.17G 0.04325 0.1071 0.004934 0.1552 1441 640: 100%|████████████████████████| 86/86 145/299 9.17G 0.04325 0.1071 0.004934 0.1552 1441 640: 100%|████████████████████████| 86/86 [00:49<00:00, 1.74it/s]
Class Images Targets P R [email protected] [email protected]:.95: 10%|█▎ | 1/10 [00:03<00:32, 3.63s/i Class Images Targets P R [email protected] [email protected]:.95: 20%|██▌ | 2/10 [00:04<00:23, 2.92s/ Class Images Targets P R [email protected] [email protected]:.95: 30%|███▉ | 3/10 [00:06<00:17, 2.55s Class Images Targets P R [email protected] [email protected]:.95: 40%|█████▏ | 4/10 [00:07<00:12, 2.0 Class Images Targets P R [email protected] [email protected]:.95: 50%|██████▌ | 5/10 [00:08<00:08, 1. Class Images Targets P R [email protected] [email protected]:.95: 60%|███████▊ | 6/10 [00:08<00:05, 1 Class Images Targets P R [email protected] [email protected]:.95: 70%|█████████ | 7/10 [00:09<00:03, Class Images Targets P R [email protected] [email protected]:.95: 80%|██████████▍ | 8/10 [00:09<00:01, Class Images Targets P R [email protected] [email protected]:.95: 90%|███████████▋ | 9/10 [00:10<00:00 Class Images Targets P R [email protected] [email protected]:.95: 100%|████████████| 10/10 [00:10<00:00 Class Images Targets P R [email protected] [email protected]:.95: 100%|████████████| 10/10 [00:10<00:00, 1.05s/it]

2.verification：
Class Images Targets P R [email protected] [email protected]:.95
all 607 7.38e+04 0.893 0.352 0.342 0.232
person 607 4.36e+03 0.846 0.798 0.776 0.536
head 607 6.42e+04 0.903 0.128 0.123 0.0635
helmet 607 5.23e+03 0.931 0.129 0.127 0.0965

from smart_construction.

sorry,I solved the problem,mainly for anchors and data labels,get the same score with author,Thank you.

你好，我也遇到了同样的问题，请问你的（yolov5s模型基础）的训练脚本的epoch是多少呢，还有你自己kmeans跑出来的anchor数值大小是和作者README一样么，非常感谢你的回答

from smart_construction.

我已更新文件 models/custom_data.yaml，更新了我经过kmeans的anchors和网络的深度和宽度，请试一下

from smart_construction.

问题已经的到解决，先关闭这个 issue 了，后续如果有疑问的话欢迎继续提 issue ，感谢大家的关注和支持！

from smart_construction.

其实这个问题是正确使用文件：./data/gen_data/merge_data.py，因为在提取人体类别和坐标的时候会生成一次人体lable，如果操作多次，会重复生成人体lable。另一个原因是其它数据库的制作人体标签时，是不能和Safety-Helmet-Wearing-Dataset数据库合并的，而是单独标签。我迁移训练一轮效果：
Class Images Targets P R [email protected] [email protected]:.95: 100%
all 3.3e+03 2.07e+04 0.516 0.858 0.815 0.47
person 3.3e+03 1.08e+04 0.408 0.811 0.729 0.44
head 3.3e+03 9.18e+03 0.48 0.862 0.814 0.345
helmet 3.3e+03 747 0.658 0.902 0.902 0.624

这里Safety-Helmet-Wearing-Dataset数据是把train和test合在一起（当然单幅图标签是独立的），大概69k，用val数据（约0.6k）测试。coco train提取的人体标签约6万多点，这样共7.1万训练集，验证集共33k。

所以，作者的测试结果是真实可信的，我没有用custom_yolov5.yaml，只用了yolov5s.yaml，当然也没有./data/gen_anchors/clauculate_anchors.py，聚类得出先验框。

from smart_construction.