[CLIFF] misc questions for evaluation
yohanshin opened this issue · comments
Dear Authors,
Hi, I am trying to reproduce the evaluation table and have a few questions.
-
I could get similar (not exactly the same though) PA-MPJPE for 3DPW dataset using HR48 backbone. But I want to be confirmed if I am doing it correctly
-- From the previous Issue post, I found that you are using the bbox obtained through projecting 3D keypoints onto image plane. In this case, which keypoints configuration did you use (e.g., COCO17, OpenPose25, H36M ...)
-- I am using the camera focal length as a known value and computed it by averaging the x and y focal lengths. Is it same as what you've done? -
Is the HR48 checkpoint uploaded in Google Drive the same as what you got for the Human36M evaluation performance?
Thank you so much for your answer in advance!
1 a). We used the COCO17 keypoint configuration.
1 b). Yes, average of f_x and f_y.
2. No. Please see #138 (comment).
Thank you for your clarifications on the first and the second questions!
For the third one, I have a follow-up question. Does it mean that the checkpoint uploaded on Google Drive will show the MPJPE of 47.6 and PA-MPJPE of 32.7?
Yes, you are right.
@yohanshin Hi! Sorry to bother you. I'm trying to reproduce the result on Human3.6M with the released model hr48-PA43.0_MJE69.0_MVE81.2_3dpw.pth
, based on the H36M Protocol 2 3D keypoint annotation and the evaluation code released by SPIN. I have successfully got the best result on 3DPW, but I just got MPJPE = 58.2
and PA-MPJPE = 39.5
, which is far away from the result given by the author. I wonder that have you reproduced the evaluation result on H36M? If so, what's your evaluation code and the test data? Thank you very much!
Hi @RammusLeo , I actually could not get the table for Human3.6M dataset. Sorry for not being so helpful..
Thank you all the same for your response! I'll try other methods.
@yohanshin Hi! Sorry to bother you. I'm trying to reproduce the result on Human3.6M with the released model
hr48-PA43.0_MJE69.0_MVE81.2_3dpw.pth
, based on the H36M Protocol 2 3D keypoint annotation and the evaluation code released by SPIN. I have successfully got the best result on 3DPW, but I just gotMPJPE = 58.2
andPA-MPJPE = 39.5
, which is far away from the result given by the author. I wonder that have you reproduced the evaluation result on H36M? If so, what's your evaluation code and the test data? Thank you very much!
hi @RammusLeo ,i got the same result as yours,I feel confused
@yohanshin Hi! Sorry to bother you. I'm trying to reproduce the result on Human3.6M with the released model
hr48-PA43.0_MJE69.0_MVE81.2_3dpw.pth
, based on the H36M Protocol 2 3D keypoint annotation and the evaluation code released by SPIN. I have successfully got the best result on 3DPW, but I just gotMPJPE = 58.2
andPA-MPJPE = 39.5
, which is far away from the result given by the author. I wonder that have you reproduced the evaluation result on H36M? If so, what's your evaluation code and the test data? Thank you very much!
are these numbers?@RammusLeo
MPJPE: 58.211485621217186
Reconstruction Error: 39.58124800903086