Fair comparison?

[yifliu3]

Dear devs,

Congratulations on your great work. I found NoPoSplat is trained and tested on the re10k dataset, while the comparison methods are trained on different datasets. So I'm wondering if this comparison is fair?

[botaoye]

Hi, thank you for your interest in our work. I am not sure what task you are referring to.

• For NVS, we trained and evaluated according to the previous pixelSplat and MVSplat settings, so it was trained on the same dataset.
• For pose estimation, we agree that we trained on RE10K, while RoMa or MASt3R were not trained on RE10K. However, this is actually one of our strengths as we do not need ground truth depth for training, so we can train our method on video datasets such as RE10K, whereas RoMa or MASt3R cannot. In addition, we also evaluated our method on the ACID and ScanNet datasets, on which we did not train. RoMa was trained on the ScanNet dataset, but we were still able to outperform it when zero-shot generalize to it.