3DPW_validation_crowd_hhrnet_result.json

[zhLawliet]

@hongsukchoi
i can‘t find the
3DPW_validation_crowd_hhrnet_result.json
J_regressor_mi_smpl.npy
MuPoTs_test_hhrnet_result.json

[hongsukchoi]

The 3dpw_validation_crowd_hhrnet_result.json is in the 2DPose_Detection folder.

I didn't prepare the MuPoTs evaluation code in this repo, since it has few crowd scenarios and is not a in-the-wild dataset.

[zhLawliet]

thanks, can you  share the  J_regressor_mi_smpl.npy or  if i  want  to  compare  the value of mupots with you in my paper ，can i  use J_regress_h36m.py  instead?

…

---Original--- From: "Hongsuk ***@***.***> Date: Thu, Nov 10, 2022 02:56 AM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [hongsukchoi/3DCrowdNet_RELEASE]3DPW_validation_crowd_hhrnet_result.json (Issue #22) The 3dpw_validation_crowd_hhrnet_result.json is in the 2DPose_Detection folder. I didn't prepare the MuPoTs evaluation code in this repo, since it has few crowd scenarios and is not a in-the-wild dataset. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[hongsukchoi]

I can share it. I uploaded to the joint_regressor folder.

[zhLawliet]

thanks

…

---Original--- From: "Hongsuk ***@***.***> Date: Thu, Nov 10, 2022 06:26 AM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [hongsukchoi/3DCrowdNet_RELEASE]3DPW_validation_crowd_hhrnet_result.json (Issue #22) I can share it. I uploaded to the joint_regressor folder. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[zhLawliet]

@hongsukchoi can you share the MuPoTs_test_hhrnet_result.json and MuPoTs_test_openpose_result.json, i want to reproduce the benchmark of MuPoTs, thanks
if i set the input_joint_name = 'gt', the pck_mean of MPJPE is 0.72376, the pck_mean of PA-MPJPE is 0.9463 ,
(pck_thresh = 150) , that's meet your expectations？

[hongsukchoi]

It takes some time to find them.

[hongsukchoi]

And if you use GT, PCK should be much higher. Does 0.72376 mean 72.376 PCK?

[zhLawliet]

yes, 72.376 PCK

…

---Original--- From: "Hongsuk ***@***.***> Date: Thu, Nov 17, 2022 02:42 AM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [hongsukchoi/3DCrowdNet_RELEASE]3DPW_validation_crowd_hhrnet_result.json (Issue #22) And if you use GT, PCK should be much higher. Does 0.72376 mean 72.376 PCK? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[hongsukchoi]

Here is the hhrnet result. I cannot find the openpose result :( But I remember using this repo: https://github.com/tensorboy/pytorch_Realtime_Multi-Person_Pose_Estimation

https://drive.google.com/drive/folders/1_Xrtd6k8sFv8FHh7NOg4B8pVBX-vyzS0?usp=sharing

[zhLawliet]

thanks

…

---Original--- From: "Hongsuk ***@***.***> Date: Tue, Nov 22, 2022 05:25 AM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [hongsukchoi/3DCrowdNet_RELEASE]3DPW_validation_crowd_hhrnet_result.json (Issue #22) Here is the hhrnet result. I cannot find the openpose result :( But I remember using this repo: https://github.com/tensorboy/pytorch_Realtime_Multi-Person_Pose_Estimation https://drive.google.com/drive/folders/1_Xrtd6k8sFv8FHh7NOg4B8pVBX-vyzS0?usp=sharing — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[hongsukchoi]

When you run openpose, there could multiple outputs in the scene or even for one person. Filter them out with this code: https://github.com/hongsukchoi/3DCrowdNet_RELEASE/blob/main/tool/match_mupots_2dpose.py

[zhLawliet]

@hongsukchoi thanks，i have done it， i find the 3DPCK of MPJPE is just 61.2. my eval code is https://github.com/ddddwee1/MuPoTS3D-Evaluation/blob/master/util/evaluate.py

[hongsukchoi]

Hi,

I think there are some bugs.

1. Use the matlab code from the original paper
2. Check the output by comparing with the input image

[zhLawliet]

ok, i try it, if i want to get the results of paper, the chekpoint is exp_04-06_23_43 epoch10 of pretrained_3DCrowdNet?

[hongsukchoi]

Yes. I haven’t checked it before the release, but it will at least give similar results. If there’s no bug and you use detected 2d poses and gt, pck will be at least over 70 and 80 respectively.

Also, compare the joint order with the MuPoTs dataset during debugging.

[zhLawliet]

yes, you are right, i can get the result. by
https://github.com/mks0601/3DMPPE_POSENET_RELEASE/blob/master/data/MuPoTS/mpii_mupots_multiperson_eval.m

[hongsukchoi]

Great! The released pre-trained weights are for reproducing 3DPW-Crowd and 3DPW, which had the fastest convergence. If you want to reproduce the exact numbers or even higher accuracy, you can train longer. Like lr_decay: [40], end_epoch: 50

[zhLawliet]

ok，thank you

…

---Original--- From: "Hongsuk ***@***.***> Date: Wed, Nov 23, 2022 01:02 AM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [hongsukchoi/3DCrowdNet_RELEASE]3DPW_validation_crowd_hhrnet_result.json (Issue #22) Great! The released pre-trained weights are for reproducing 3DPW-Crowd and 3DPW, which had the fastest convergence. If you want to reproduce the exact numbers or even higher accuracy, you can train longer. Like lr_decay: [40], end_epoch: 50 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>