search

hongsukchoi / Pose2Mesh_RELEASE

Official Pytorch implementation of "Pose2Mesh: Graph Convolutional Network for 3D Human Pose and Mesh Recovery from a 2D Human Pose", ECCV 2020
MIT License
671 stars 69 forks source link

visualization - 2D keypoints with input image #63

Closed moonsh closed 1 year ago

moonsh commented 1 year ago

Thank you for sharing your work.

I just wanted to visualize the input image with 2d pose estimation using the 3DPW test data set to see if the 3d estimation is correct.

I found the below code for visualization but strangely the 2d pose and image are severely misaligned. This code is not available? https://github.com/hongsukchoi/Pose2Mesh_RELEASE/blob/master/data/PW3D/dataset.py#L216

hongsukchoi commented 1 year ago

Hi,

There are some things to check

  1. Is the gt 2d pose aligned with the image?
  2. Does the scale of data['pred_pose2d'] match the image's width and height?

I think the visualization code in get_item should work ok, could you provide your visualization?

moonsh commented 1 year ago

This command I used for testing "python main/test.py --cfg asset/yaml/posenet_cocoJ_test_3dpw.yml"

One example looks like this. The 2D pose and image are not aligned. However, I checked the 3D pose looks fine.

1

and also I found that the evaluation result is strange. It seems like something is wrong.

MPJPE: 263.8583 Evaluation start... COCO MPJPE (mm) >> tot: 263.87 COCO PA-MPJPE (mm) >> tot: 114.80

hongsukchoi commented 1 year ago

Hmm, I recently changed the 3dpw annotation files and maybe that's the problem.

moonsh commented 1 year ago

I see. Can you share the previous one?

moonsh commented 1 year ago

I got the previous annotation file here. https://github.com/kasvii/GATOR/issues/2

hongsukchoi commented 1 year ago

@moonsh Sorry for the late reply. I updated the annotation files and checked the visualization. And thank you for letting me know.