Issues with using 3D Models in co3d Code

[cchang021]

Hello, I really appreciate your exceptional work! It has been highly useful for me. I observed that in your co3d code (without optical flow), the dataset comprises a sequence of teddy bear images. I'm looking to apply the code to my dataset, which contains images of a 3D model from some various views, almost same to the 'cow' data in the shape_from_silhouette code except the particular model.

However, when attempting to just replace the bear images and masks with pictures of the 3D model, I encountered an issue where the metaball struggles to estimate depth and color for the 3D model data. Despite the loss decreasing, it remains significantly high after optimization, and the model fails to predict shape, depth, or color.

Could you kindly share if you've had the opportunity to explore the use of 3D models as a dataset for the co3d code, or if you've come across a similar challenge?