if using train_depth only, what's difference between MVSFormer or other mvsnet

[liucsg1128]

No description provided.

[haofeixu]

Thank you for the great question! In our paper, we already compared our approach with MVSFormer's design in Table 2 (single branch), where our method significantly outperforms it. However, beyond this, there are other notable differences between our model and prior MVS networks.

Most state-of-the-art MVS models are trained with ground truth depth supervision and typically evaluated on closed test sets such as DTU or Tanks and Temples. While these models achieve impressive results on popular benchmarks, they often struggle to generalize to in-the-wild images in an open domain. This limitation hampers their practical use in diverse, real-world applications.

To address this gap, we introduce Gaussian splatting as an unsupervised pre-training objective, which enables us to train more robust and generalizable MVS models on large-scale unlabelled datasets. As demonstrated on our project page (Scale-Consistent Depth Prediction), our model shows strong generalization to unseen, in-the-wild samples, an area where existing MVS models often fall short.

In summary, our method not only incorporates technical innovations but also focuses on delivering a generalizable and robust MVS solution for open-world scenarios, setting it apart from prior approaches.