Closed xiyichen closed 2 months ago
The default parameter setting is for FLAME 2017. Maybe you can simply try to modify the weight of regular loss to get better results.
Increasing both regularization weights to 1e-5 indeed solves the problem. Thank you!
Thanks for the work and code release! I tested it on a custom multi-view dataset with ground truth camera parameters. Using the BFM model, I am able to get desired results, but using FLAME 2020 (with the same camera parameters and landmarks), the resulted mesh looks very distorted. BFM result:
FLAME result:
![vis_1_flame](https://github.com/YuelangX/Multiview-3DMM-Fitting/assets/50854238/6324ca00-9b2a-4139-9403-3ba7aae96433)
I wonder if you used a different FLAME version. I also tried 2019 but it's also not working.