search

mks0601 / 3DMPPE_ROOTNET_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
477 stars 65 forks source link

test on wild images #36

Closed dldaisy closed 2 years ago

dldaisy commented 2 years ago

Hi, I want to get the root depth in ndc space from a wild image, but the image was captured by another camera with different intrinsics(the focal length is different from 1500 used in the code). Can I run rootnet inference on the image? If I can, how can I get the correct depth in NDC space？ Thanks a lot!

mks0601 commented 2 years ago

RootNet predicts root depths in the focal length-normalized space. I set the normalized focal lengths as (1500,1500), so the predicted depth is in (1500,1500) focal length space. If you know true focal length, you can convert the predicted depth to the correct focal length space.

dldaisy commented 2 years ago

Hi @mks0601 Thanks for reply! Does it mean I don't need to retrain the model but only needs to modify the focal length in demo.py?

mks0601 commented 2 years ago

Yes.

dldaisy commented 2 years ago

I see. Thanks a lot!

nicolasugrinovic commented 2 years ago

Hi @mks0601, Kudos for the great work. I see that during the evaluation for MuPots dataset, you read this focal length parameter from MuPoTS-3D.json file. I was wondering, how did you get instrinsics parameters? Did you calculate these values or they come from the dataset. I believe I could not find them in the dataset. Thanks

mks0601 commented 2 years ago

I optimized the focal lengths by fitting GT 3D joint coordinates to GT 2D joint coordinates

nicolasugrinovic commented 2 years ago

Thanks for the reply, I really appreciate it!