About Splatfacto's color correct method #3737
Description
While experimenting with various options during the project, I accidentally enabled the color_corrected_metrics option. This option significantly slowed down the training process. Of course, this was entirely my fault, as the default value is False. The few hours I spent investigating the slowdown were simply the cost of that mistake.
That said, it feels problematic that this option can be enabled during training, rather than being restricted to evaluation. While color correction is a reasonable approach for fair metric evaluation, computing it over millions (or even tens of millions) of Gaussians during training seems unnecessary and inefficient.
When color_corrected_metrics is enabled, the torch.linalg.lstsq operation becomes a major bottleneck due to its high computational cost, and this overhead is incurred at every training iteration.
I believe the training pipeline should be modified so that color-corrected metrics (e.g., cc_psnr) are not computed during training, or alternatively, that the color correction is performed only on a sampled subset rather than the full set of Gaussians.