search

mks0601 / 3DMPPE_POSENET_RELEASE

Official PyTorch implementation of "Camera Distance-aware Top-down Approach for 3D Multi-person Pose Estimation from a Single RGB Image", ICCV 2019
MIT License
807 stars 147 forks source link

other public datasets，overfitting issue #135

Open YHaooo-4508 opened 11 months ago

YHaooo-4508 commented 11 months ago

Hello, I am trying to adapt the MPII_3DHP dataset on your code, but it's totally overfitting First I train it on H36M+MPII，the MPJPE is about 50+. This is normal Then i change the dataset using MPII_3DHP+MPII to train. 50MPJPE in trainset but 200+ in validate dataset. Have you conducted experiments on MPII_3DHP dataset？ Why is there such a serious overfitting？

mks0601 commented 11 months ago

I'm not clear about your experiments.

  1. First I train it on H36M+MPII，the MPJPE is about 50+. This is normal -> what is the test set?
  2. Then i change the dataset using MPII_3DHP+MPII to train. 50MPJPE in trainset but 200+ in validate dataset. -> what are the training and validation sets?
YHaooo-4508 commented 11 months ago

I'm not clear about your experiments.

  1. First I train it on H36M+MPII，the MPJPE is about 50+. This is normal -> what is the test set?
  2. Then i change the dataset using MPII_3DHP+MPII to train. 50MPJPE in trainset but 200+ in validate dataset. -> what are the training and validation sets?

1、train with H36M + MPII2D train dataset；test with H36M validate dataset 2、train with MPII_3DHP+MPII2D train dataset；test with MPII_3DHP validate dataset

mks0601 commented 11 months ago

I don't think this is from overfitting. I guess there is something wrong in your evaluation code?

YHaooo-4508 commented 11 months ago

I don't think this is from overfitting. I guess there is something wrong in your evaluation code?

I checked the code and found that there was an issue with bbox, which has been fixed. However, the results of subsequent testing were still overfitting. The validation set MPJPE was around 150, trainset is about 60. Here are some of my visualization results. Some samples also have poor results image image image image

mks0601 commented 11 months ago

Error 60 on training set means definitely underfitting

YHaooo-4508 commented 11 months ago

Error 60 on training set means definitely underfitting

Firstly, thank you for communicating with me. I know that continuing training may improve performance in the training set, but at this point, the validation set error no longer decreases, so I stopped training, and there was a significant difference between the validation set error and the training set error. By the way, during my training and validation process, I did not use object detection algorithms to predict bbox, but instead used key point positions to fit the bounding box. I am not sure if this method will affect the training.