validation by my own 2d bbox detector.

[lxy5513]

First of all , thank for you awesome project.

For to use the pose estimator on real-time, I insert YOLOv3 to get human bbox, now I want test the AP accuracy of human-pose-estimation with new human detector's pose.

but when I set the confidence of YOLOv3 to 0.0, I get 687810 bboxes, like this

=> num_images: 5000
=> Total boxes: 43818
=> Total boxes after fliter low [email protected]: 687810
=> load 687810 samples

when set confidence to 0.1 , like this

Arch	AP	Ap .5	AP .75	AP (M)	AP (L)	AR	AR .5	AR .75	AR (M)	AR (L)
384x288_pose_resnet_50_d256d256d256	0.107	0.137	0.112	0.047	0.192	0.113	0.145	0.116	0.050	0.199

=> num_images: 5000
=> Total boxes: 16574
=> Total boxes after fliter low [email protected]: 16574
=> load 16574 samples

The AP 
| Arch | AP | Ap .5 | AP .75 | AP (M) | AP (L) | AR | AR .5 | AR .75 | AR (M) | AR (L) |
|---|---|---|---|---|---|---|---|---|---|---|
| 384x288_pose_resnet_50_d256d256d256 | 0.166 | 0.212 | 0.178 | 0.089 | 0.271 | 0.172 | 0.214 | 0.181 | 0.093 | 0.281 |

I can't understand why you can get 104125 bboxes, how do you set the threconfidence?

=> num_images: 5000
=> Total boxes: 104125
=> Total boxes after fliter low [email protected]: 104125
=> load 104125 samples

could you please help me ? Thanks in advance.