Question about joint numbers

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

708 stars 130 forks source link

Question about joint numbers #128

Open PhysShaw opened 7 months ago

PhysShaw commented 7 months ago

Hi! Mr.Gyeongsik, Thanks for your excellent work and code!May I ask a question that in Demo.py there are 29 joint numbers,while the trained dataset Human3.6M and MSCOCO have different joint numbers.I mean,is it possible to use 17 or 18 as joint numbers(MSCOCO format) in the demo.py?

mks0601 commented 7 months ago

Hi, you can extract 17/18 joints out of 29 joints by checking this joints_name https://github.com/mks0601/I2L-MeshNet_RELEASE/blob/eae673ea0b5229c6c7c0ed6d24633f648683da5c/demo/demo.py#L54

PhysShaw commented 7 months ago

Thank you for your prompt answer!

I'm using your snapshot_7.pth as ckpt file right now,it seems that it's used for 29 joint numbers.How can I change joint numbers with snapshot_7.pth as ckpt file?
Or it means that I should train a new ckpt file with 17 joint numbers?

mks0601 commented 7 months ago

You don't have to retrain it. You can just extract your 17 out of predicted 29 joints.