Parameter finetuning vs Output finetuning

[carsonswope]

It seems that running gradient descent for the depth prediction network makes up the majority of the runtime of this method. The current MiDaS implementation (v3?) contains 1.3 GB of parameters, most of which are for the DPT-Large (https://github.com/isl-org/DPT) backbone.

In your research, did you experiment with performance differences between 'parameter finetuning' and just simple 'output finetuning' for the depth predictions (like as discussed in the GLNet paper (https://arxiv.org/pdf/1907.05820.pdf))?

I would also be curious about whether as a middle ground, maybe just finetuning the 'head' of the MiDaS network would be sufficient, and leave the much larger set of backbone parameters locked.

Thanks!