Do we need ground truth depth to train the model ?
Dear author,
Do we need ground truth depth to train the model? It is easy to obtain coarse depth , but not easy to obtain ground truth depth for real world human 3d reconstruction .
In your paper, you mentioned that , the learning target, or the residual diparity, is calculated by the ground truth disparity maps and the coarse disparity maps. Therefore, I am very confused. In real world case, it is not easy to obtain the ground truth disparity.
Thanks so much!
Dear Jiakui,
We use 3D scan dataset such as THuman2.0 to generate the ground truth disparity maps and train our model. You can apply for the THuman2.0 dataset which is public for academic use. Though our diffusion model is only trained on the synthetic dataset, it performs well in real-world cases.