search

Vegetebird / StridedTransformer-Pose3D

[TMM 2022] Exploiting Temporal Contexts with Strided Transformer for 3D Human Pose Estimation
MIT License
335 stars 37 forks source link

How to use ground truth 2D joint locations as input in the code? #38

Closed zhang-yuxian closed 7 months ago

zhang-yuxian commented 10 months ago

How to use ground truth 2D joint locations as input in the code?

屏幕截图 2023-12-14 210017
Vegetebird commented 10 months ago

Hi~ You can use the command: --keypoints 'gt'

zhang-yuxian commented 10 months ago

Thank you! I have successfully solved this problem.

---- Replied Message ---- | From | Wenhao @.> | | Date | 12/15/2023 21:15 | | To | Vegetebird/StridedTransformer-Pose3D @.> | | Cc | zhang-yuxian @.>, Author @.> | | Subject | Re: [Vegetebird/StridedTransformer-Pose3D] How to use ground truth 2D joint locations as input in the code? (Issue #38) |

Hi~ You can use the command: --keypoints 'gt'

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

cqs0925 commented 8 months ago

Hey, thanks for your code. I've tried using the pretrained model. Fine-tuned CPN gives the exact same result, but ground truth input only gives 37.31/30.03(Protocol 1/2) while in the paper it is 28.5 under protocol 1. I've tried to test without refine module and cpn give even better result than gt. I've tried this code twice and it gives the same problem. I am wondering is there another pretrained model that are used to process GT input.

zhang-yuxian commented 8 months ago

For ground truth as input, change "self.parser.add_argument('-k', '--keypoints', default='cpn_ft_h36m_dbb', type=str)" to "self.parser.add_argument('-k', '--keypoints', default='gt', type=str)", then train and test. If you get worse results than this paper, please check "frames(351)" and "batch_size". Actually, Increasing “batch_size” value can reduce MPJPE, like “batch_size=280”/“batch_size=300”....but it requires larger graphics memory！Please try again!

在 2024-01-24 20:12:50，"cqs0925" @.***> 写道：

Hey, thanks for your code. I've tried using the pretrained model. Fine-tuned CPN gives the exact same result, but ground truth input only gives 37.31/30.03(Protocol 1/2) while in the paper it is 28.5 under protocol 1. I've tried to test without refine module and cpn give even better result than gt. I've tried this code twice and it gives the same problem. I am wondering is there another pretrained model that are used to process GT input.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>