Closed zwl995 closed 1 month ago
This dataset contains images in various lighting conditions. The model simply adds floaters to overfit such lighting differences.
Hi, @zwl995, during the development of our CityGaussian, the influence of light changes is ignored and thus can lead to undesirable floaters. To remove them, a favorable practice is introducing appearance embedding like VastGaussian.
Hello, I used the configuration file you provided to train the rubble_c9_r4 scenario in the mill19 dataset, but the result I got was more clutter in the air. Can you help me see what caused it?