search

YuliangXiu / ECON

[CVPR'23, Highlight] ECON: Explicit Clothed humans Optimized via Normal integration
https://xiuyuliang.cn/econ
Other
1.08k stars 106 forks source link

bni.obj reconstruct failed #33

Open surheaven opened 1 year ago

surheaven commented 1 year ago

image

Can you tell me what caused this to happen? thanks

surheaven commented 1 year ago

When I use the Perspective Camera, it again causes the following problems image

xucao-42 commented 1 year ago

From the first image, it looks like the mask is not provided correctly. Would you share the input image and the estimated normal maps? Also, could you provide more details about how you use a perspective camera? ECON should assume an orthographic camera for the input image.

surheaven commented 1 year ago

image & normal: 000000 normal and i use perspective camera to get the depth map: depth while the depth in orthographic camera: depth_orth I modified it here to get the depth chart: image

surheaven commented 1 year ago

And i also use the code from bilateral_normal_integration to get z in perspective camera. image But when do np.log() it failed because the depth is -25856. So can you tell me how to modify in perspective camera? And i find the depth is processed as followed, so how can i process depth in perspective camera? Thanks! image

xucao-42 commented 1 year ago

If you borrow the code from bilateral_normal_integraion, please ensure you have provided the mask, otherwise the entire image will be integrated (in you case the black backgroud). Also, please check this part for handling persepective normal maps is included in your code. https://github.com/xucao-42/bilateral_normal_integration/blob/fd853ba2cbb59be6a9e09f9418523a85c7ba4cf0/bilateral_normal_integration_cupy.py#L196

You can try perspective camera, but it should only work well for the front view but not for the back view. I am not so sure what's going to happen for the back view (maybe large distortion).