search

xuxiaoxxxx / 3DSS-VLG

1 stars 0 forks source link

A discussion about the paper #2

Closed Rickilous closed 3 weeks ago

Rickilous commented 3 weeks ago

Hi,

I recently came across your article and found it very inspiring. However, I wanted to ask if you've noticed the growing trend in 3D open vocabulary semantic scene understanding. These works have progressed to a point where 3D labels are no longer necessary.😮

I'm currently working on weakly supervised semantic segmentation and feel a bit confused — in comparison to these studies that don't require 3D labels, is using 2D models for weakly supervised learning already outdated?😣 I'd love to hear your thoughts on this.

Would it be possible to have a private discussion? You can reach me at my email, which is listed on my homepage: jiangyu_cdut59@163.com.😀

xuxiaoxxxx commented 3 weeks ago

You're right, open-vocabulary segmentation tasks without 3D labels are gaining popularity. However, it's important to note that these tasks still depend on additional 2D pre-trained models, such as CLIP, SAM, etc., as they provide crucial prior knowledge to guide the 3D semantic segmentation models. Without this guidance, the models would struggle to perform effectively. So, in my view, using 2D pre-trained models to support 3D semantic segmentation in low-label tasks—like weakly supervised or zero-shot segmentation—remains a key trend.

Rickilous commented 3 weeks ago

I appreciate your valuable insights. I hope your paper can be accepted soon. Good luck. ♥