Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

mAP difference in MSCOCO val #190

Closed
zw-xxx opened this issue Jun 23, 2020 · 3 comments
Closed

mAP difference in MSCOCO val #190

zw-xxx opened this issue Jun 23, 2020 · 3 comments

Comments

@zw-xxx
Copy link

zw-xxx commented Jun 23, 2020

您好!HRNet的工作真的非常棒,非常感谢开源代码!
有几个问题,想请教一下:
person_keypoints_val2017.json和COCO_val2017_detections_AP_H_56_person.json有什么区别吗?
如果理解无误的话,使用的都是val2017验证集。那为什么在每次训练epoch结束的模型评价mAP值,要比test.py所得的mAP 低

这是在train中最后一个训练轮次结束所得mAP

2020-06-22 19:56:58,671 | Arch | AP | Ap .5 | AP .75 | AP (M) | AP (L) | AR | AR .5 | AR .75 | AR (M) | AR (L) |
2020-06-22 19:56:58,672 |---|---|---|---|---|---|---|---|---|---|---|
2020-06-22 19:56:58,672 | pose_hrnet | 0.744 | 0.926 | 0.816 | 0.722 | 0.783 | 0.774 | 0.935 | 0.838 | 0.746 | 0.817 |

这是通过val检测final_state.pth的mAP

2020-06-23 04:13:00,041 | Arch | AP | Ap .5 | AP .75 | AP (M) | AP (L) | AR | AR .5 | AR .75 | AR (M) | AR (L) |
2020-06-23 04:13:00,043 |---|---|---|---|---|---|---|---|---|---|---|
2020-06-23 04:13:00,043 | pose_hrnet | 0.724 | 0.892 | 0.799 | 0.692 | 0.789 | 0.780 | 0.932 | 0.848 | 0.740 | 0.839 |

两者相差了2个点,我认为这应该与json文件的选取相关。
以及,请问在论文中:Table 1. Comparisons on the COCO validation set,使用的是哪个mAP值。
还望予以赐教,谢谢!

@leoxiaobin
Copy link
Owner

person_keypoints_val2017.json is the ground truth for dataset coco val2017.
COCO_val2017_detections_AP_H_56_person.json provides the person bounding boxes from a person detector.

@zw-xxx
Copy link
Author

zw-xxx commented Aug 23, 2020

thank u very much!

@zw-xxx zw-xxx closed this as completed Aug 23, 2020
@chenhaomingbob
Copy link

在coco 验证集的精度上我也遇到了相似问题。
我使用 COCO_val2017_detections_AP_H_56_person.json作为bbox文件,并且加载了 pose_hrnet_w48_384x288.pth 模型。得到了以下的精度

Model AP Ap .5 AP .75 AP (M) AP (L) AR AR .5 AR .75 AR (M) AR (L)
pose_hrnet 0.7411 0.8968 0.805 0.696 0.8168 0.7866 0.9268 0.843 0.7394 0.8548

我测试的AP值为0.741,未达到论文中报告的0.763。

除了用检测的bbox进行测试,我还使用了来自person_keypoints_val2017.json的 gt bbox。以下是我的测试结果:

Model AP Ap .5 AP .75 AP (M) AP (L) AR AR .5 AR .75 AR (M) AR (L)
pose_hrnet 0.762 0.9253 0.8239 0.7269 0.8193 0.7906 0.9399 0.8449 0.7516 0.8493

与论文中的结果0.763不符。

您能解决我的疑惑吗?十分感谢!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants