Closed CscotfordMV closed 3 weeks ago
Since the code of the Gaussian renderer is copied from 3D Gaussian Splatting, I cropped the image and modified the intrinsic in lib.dataset.Dataset.GaussianDataset, so that the camera principal point is at the center of the image. I plan to remove this unnecessary operation in the next few days.
@CscotfordMV How did you extract the camera parameters of your custom data? When i train on my custom data, with camera parameters extracted via reality capture, the train_meshhead step fails and throws the error: Calculated padded input size per channel: (0). Kernel size: (1). Kernel size can't be greater than actual input size
All my input sizes are the same as in the demo dataset, so i thought it may be a problem with the camera parameters, that cause problems downstream. When i plot my vertices from subject/params/*/vertices.npy i can see their orientation being wrong compared to the demo data
@NikoBele1 I use the camera parameters provided by the NeRSemble data. I guess they are calibrated using checkerboard by the opencv function. If the camera parameters are wrong, it may be caused by inconsistent coordinate systems.
I've managed to train the MeshHead on some custom data and it worked really well, results looked great. However, when I try and train the GaussianHead model afterwards, it seems all the results saved (and I presume the images it's loading in) are being saved with a large amount of padding around the top and left side so the image is shifted a great deal being replaced with just black. I believe I have edited the configs correctly (very little actually as I started with the same resolution size as the mini_demo_dataset - 2080x2080). I should add too that the code works perfectly well with the mini-demo-dataset you provide so the issue isn't any changes I have made myself. Where is this padding coming from?