search

CyberAgentAILab / SuperNormal

[CVPR 2024] Official implementation of "SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration"
MIT License
145 stars 7 forks source link

How to do it on custom data? #7

Closed YiChenCityU closed 1 month ago

YiChenCityU commented 5 months ago

Hi, Could you provide some information about reconstruct on custom data? And how to get normal maps? Thanks very much.

flow-specter commented 3 months ago

Same question

xucao-42 commented 3 months ago

Hi, To obtain the normal maps, follow the instructions in this CVPR'23 work: https://github.com/satoshi-ikehata/SDM-UniPS-CVPR2023 In short, at each viewpoint, you need to capture a set of images under varying illumination (from my experience, about 10 images per view is sufficient). This can be realized by moving a handheld light source. Input the images of each view to the CVPR'23 work should output a good quality normal map.

Repeat this process for each view, and you can get multi-view normal maps. Since you also captured rgb images, pick one from each view to calibrate the camera parameters. We found Colmap usually fails for such a capture setup, but MetaShape performs more robustly.

We recommend placing the object on a turntable, but moving the camera and tripod should also be fine. Ensure there is a textured background static to the target object or use markers to increase the success rate of camera calibration.

xucao-42 commented 1 month ago

Instructions on preparing custom data is available here.