search

muelea / shapy

CVPR 2022 - Official code repository for the paper: Accurate 3D Body Shape Regression using Metric and Semantic Attributes.
https://shapy.is.tue.mpg.de/
319 stars 47 forks source link

how to test own dataset on Demo.py #26

Open TharHtetAungg opened 1 year ago

TharHtetAungg commented 1 year ago

Hello, I want to try my own dataset to test regressor. Which process that I need to do first before I run on regressor? Do I need to run openpose and get json format first? And also image size is only 300 450 px ?

I run my own image and got this error : ValueError: all the input array dimensions for the concatenation axis must match exactly, but along dimension 1, the array at index 0 has size 135 and the array at index 1 has size 131

My image is :1920 1080 px but I resize to 300 450 px to match with yours and still have that error.

neoglez commented 1 year ago

I have the same problem. It seems that it is not possible to input an in-the-wild image to the network before adding to the target a significant amount of information such as the image can be transformed before it goes into the network.

muelea commented 1 year ago

We use OpenPose keypoints to get a person's bounding box which we use to center and crop them. If you have the bounding box already, you can skip this step. To do this, in the data loader, replace kpts_to_bbox() with the bounding box information [xmin, ymin, xmax, ymax]. I still recommend running OpenPose and use the automated bounding box computation, because these bounding boxes will match the ones the model was trained with. You can follow the images and openpose structure in our demo and then run the SHAPY demo.

neoglez commented 1 year ago

@muelea Thanks for the helpful information. I will attempt to make a PR with a notebook demonstrating the complete pipeline.

TharHtetAungg commented 1 year ago

@muelea Hi, can I ask about which part of your code can change to T pose ? Could you please show me?

felixshing commented 1 year ago

@muelea Thanks for the helpful information. I will attempt to make a PR with a notebook demonstrating the complete pipeline.

@neoglez Hey I would like to ask have you successfully done this pipeline that tests own dataset?

neoglez commented 10 months ago

@felixshing Ja, I did but I had to make a lot of changes. I'm still trying to double check because the results I obtained are really strange.

rshtern commented 10 months ago

Can you explain how to train from an image without openpose?