About benchmark results in those sparse reconstruction methods

[Maple-geekZhu]

Firstly, thank you for your excellent work!
I have the following questions: How were the results in the appendix obtained? I noticed that a method like FSGS is used for reconstructing the entire scene, whereas in your experimental setup, it should be object-level. When you are evaluating the metrics, do you calculate them based on the ground truth and rendered images superimposed with masks (in other words, assessing the quality of the image only within the masked areas)

[chensjtu]

Yes, we follow the setting in zerorf, directly using the whole image to eval. Some extra details can be found in Appendix A.3. Since some objects are pretty small in pixel space, this eval may lead to higher performance since the background are same. We have considered only using the object pixel for eval, but this leads to problem with SSIM and LPIPS since the require image input.