search

mks0601 / I2L-MeshNet_RELEASE

Official PyTorch implementation of "I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image", ECCV 2020
MIT License
710 stars 130 forks source link

Some thoughts on improving the model #59

Open booker-max opened 3 years ago

booker-max commented 3 years ago

1.Hello, improving the resolution of the heatmap is helpful to the improvement of the indicators of MPJPE and PA-MPJPE. Have you tried to further increase the resolution of the heatmap to 128128,128, or even higher 256256,256? Is it necessary for me to do such an operation and then retrain the model? Although I think the effect of snapshot12.pth is quite good.

  1. I used snapshot12.pth and found that the effect is already very good, but all ten fingers are straight. Do you have any good suggestions for improving hand posture？
mks0601 commented 3 years ago
  1. Increasing to 128 or 256 marginally affected the performance.
  2. I don't know what you meant by snapshot_12, but you'd better give me some image results.
booker-max commented 3 years ago

1.Okay, does this mean that the performance improvement is small? 2.I use the pre-trained model snapshot_12.pth.tar, the effect is like this image image this is pretty good, but the ten fingers are straight, and I want to improve it further, do you have any good suggestions for improving hand posture？

mks0601 commented 3 years ago
  1. Yes
  2. Most of current 3D pose/shape estimation methods, including I2L-MeshNet, are designed only for the body-only case. My recent work addressed this problem. It will be updated in 4/2, so you'd better read this after the update. link: https://arxiv.org/abs/2011.11534
booker-max commented 3 years ago

Okay, thank you, I will read this article carefully.