Good Evening,
Thank you for this great work. However, I wanted to enquire about how you mask the MSE loss. The paper mentions using alpha to mask create masks but I didn't get it, since alpha refers to the confidence hyperparameter. If it referes to the confidence that is predicted by the point head, then how would you use it to mask the photometric loss? Since in this case these predicted confidence values are associated with the gaussian splats and not rendered/target pixels. Could you please elaborate more about it?
Also, do you use the "context" views for photometric supervision or only the N-Extra?
Good Evening,
Thank you for this great work. However, I wanted to enquire about how you mask the MSE loss. The paper mentions using alpha to mask create masks but I didn't get it, since alpha refers to the confidence hyperparameter. If it referes to the confidence that is predicted by the point head, then how would you use it to mask the photometric loss? Since in this case these predicted confidence values are associated with the gaussian splats and not rendered/target pixels. Could you please elaborate more about it?
Also, do you use the "context" views for photometric supervision or only the N-Extra?