Question about the choice of decoder head in MoGe / MoGe-2

[ZhouZhanbo]

Hi, thank you for releasing this excellent work and the codebase.

I have a question about the decoder design in MoGe and MoGe-2. From my understanding of the implementation, both models do not use a DPT-style decoder head. Instead, they adopt a custom lightweight convolutional head for dense prediction. Roughly speaking, this head seems to have significantly fewer parameters than a standard DPT head, around one third in my estimation.

Since DPT-style heads are widely used in monocular depth estimation and dense prediction tasks, I was wondering about the motivation behind this design choice.

Specifically, I would like to ask:

1. Was the lightweight convolutional head chosen mainly for efficiency, such as reducing parameters, memory usage, or inference cost?
2. Did you observe any empirical advantage of this head over a DPT-style head for point-map prediction or normal prediction?
3. Is the DPT head less suitable for MoGe because the model predicts 3D point maps / metric geometry rather than only depth?
4. Were there any internal ablation experiments comparing the proposed head with a DPT-style decoder head, even if they were not included in the paper?
5. In your experience, would replacing the current decoder with a DPT-style head be likely to improve accuracy, or would it bring limited benefit compared with its extra computational cost?

I am asking because I am interested in extending MoGe-style point-map prediction to video geometric dense prediction, and the decoder-head design seems important for balancing accuracy and efficiency.

Thanks again for your great work!