This repo contains the projects: 'Virtual Normal', 'DiverseDepth', and '3D Scene Shape'. They aim to solve the monocular depth estimation, 3D scene reconstruction from single image problems.
I'm a computer vision student, i'm actually doing a final year research project on the subject of 3D reconstruction based on a single image, I find your work astonishing, i congratulate your for that.
In this pull request, i have fixed some compatibility issues with the latest version of torchsparse (1.4 i guess), and also the latest version of pytorch and cuda11.7.
I have seen in one of the issues answers that the training code for the PCMs aren't released because of IP issues, it is completely understandable but i wonder if you might share with me how did you procede, with regard to the data generation process, the magnitude of the perturbations applied to the focals and the depth, what cost function have you used and any idea or intuition about the hyper parameters.
Hello,
I'm a computer vision student, i'm actually doing a final year research project on the subject of 3D reconstruction based on a single image, I find your work astonishing, i congratulate your for that.
In this pull request, i have fixed some compatibility issues with the latest version of torchsparse (1.4 i guess), and also the latest version of pytorch and cuda11.7.
I have seen in one of the issues answers that the training code for the PCMs aren't released because of IP issues, it is completely understandable but i wonder if you might share with me how did you procede, with regard to the data generation process, the magnitude of the perturbations applied to the focals and the depth, what cost function have you used and any idea or intuition about the hyper parameters.
Your answer would be very much appreciated.