0
$\begingroup$

I've been learning to do some depth estimation tasks so I came across the GLP model (https://arxiv.org/pdf/2201.07436.pdf) which is suggested on huggingface website. I'm new to machine learning, so please excuse my basic questions.

My question is: what are the main differences (conceptually) between the GLP model introduced in this paper and the previous DPT model (https://arxiv.org/pdf/2103.13413.pdf)? The performance of GLP model appeared to be better than DPT ("...our model achieves higher performance than the recently developed state-of-the-art models (Adabins, DPT) with lesser parameters...") but what are the main innovations which contribute to this improvement?

Thank you!

Here's the architecture diagrams for the two models for comparison.

GLP: enter image description here DPT: enter image description here

  • Both GLP and DPT are vision transformers (ViT) based. ViT was used for the encoder instead of CNN in both models. However, the GLP paper claimed that: "DPT use CNN-based encoders and transformers simultaneously which increases the computational complexity...In contrast to these studies, our method use only one encoder", but I don't really see from the DPT diagram that CNN encoder were used anywhere more than in GLP. Both models embedded the input image into patches and pass them into a series of transformers, from what I can see. So, is GLP model supposed to be a more 'pure' implementation of ViT than DPT model, and in what ways?

  • The 'Global-to-Local-Path' seems to be one main feature of the GLP model (as it is in the name). But while GLP model has SFF modules to (presumably) fuse global and local features, the DPT model also have the Fusion modules which seem to do a similar thing. Is the contribution of GLP model the skip connections in the SFF?

  • One thing that I can appreciate in GLP paper is the experimentation with vertical Cut Depth data augmentation. I think this is not discussed in the DPT paper (since the focus was on a more general tasks rather than just depth estimation as in GLP paper).

$\endgroup$

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.