search

tsunghan-wu / Depth-Completion

🍀 Official pytorch implementation of "Indoor Depth Completion with Boundary Consistency and Self-Attention. Huang et al. RLQ@ICCV 2019."
https://arxiv.org/abs/1908.08344
MIT License
182 stars 18 forks source link

Why not directly extract boundary of estimated depth by Sobel filter #26

Closed JCATCV closed 1 year ago

JCATCV commented 1 year ago

Hi,

I see in this work a boundary consistency network is proposed to predict boundary, but I wonder why not extracting boundary directly by the Sobel filter. This is more intuitive and much simpler, right? Are there some explanations?

thank you for your instructive work.

tsunghan-wu commented 1 year ago

Hi @JCATCV,

Thanks for asking a great question. The idea of adding boundary estimation module is to regularize the produced depth. As you can see in Fig. 1 and 5, without the module, the boundary is blur and not satisfied.

To address the issue, we first extract the ground truth depth boundary by applying Sobel filter to the ground truth depth. Then during the training, the depth completion network along with the boundary consistency network is trained end-to-end to ensure our generated depth can also produce sharp edges.

I hope the above explanations do address your concerns. Let me know if you have further questions!

Best, Tsung-Han

JCATCV commented 1 year ago

@tsunghan-wu

Thank you for reply.

So, you mean this network is mainly to facilitate training not just extracting depth boundary.

Thanks.

tsunghan-wu commented 1 year ago

@JCATCV,

Correct ;)