unsupervised-depth-completion-visual-inertial-odometry icon indicating copy to clipboard operation
unsupervised-depth-completion-visual-inertial-odometry copied to clipboard

Published 20 hours ago • verified publisher icon alexklwong

The request of voiced-VOID150 and KITTI pretrained model or bash scripts

Open myboyhood opened this issue 2 years ago • 2 comments

Hi Alex Wong! This is a really great work for developers in robotics.

I am doing the work to recover the sparse SLAM point to dense points for obstacle map construction. The scanffolding idea is brilliant. I originally plan to use Delaunay mesh map to construct obstacle map. But using your work, I will get a more precise and smooth map. That is promising and exciting.

  1. Question1

    I use OpenVINS to generate about 100 points per frame, thus I think voiced-void150.pth may be more suitable than the released pretrained vocied-void1500.pth. I noticed your response in closed issue. Unfortunately, I am new to this field. It is difficult for me to tune the hyperparameters in training void150 model.

    Since I have carefully read your RAL paper and data comparsion using void150. I hopes to request a already pretrained model or a training bash scripts for guidance. I want to try it performance first and then train the dataset by myself.

  2. Question2

    I also aims to complete depth completion at far distance, such as 10m~25m. I think the pretrained model from KITTI dataset is more suitable for this scene. Could you help to upload this pretrained model too ? I want to see its performance and I would collect data myself to train a new model.

Thanks for you any help ~

Best.

myboyhood avatar Sep 27 '23 03:09 myboyhood

Hi, yes we do have plans to do this. While I am a firm believer in continuing to support and serve the community with this work, our group does have other ongoing projects that are taking up the majority of lab compute... so things may go a bit slow for new releases. A promising work to try is KBNet: https://github.com/alexklwong/calibrated-backprojection-network

KBNet typically can handle sparser inputs than VOICED, and also has KITTI pretrained weights available. Nonetheless, I will try to release more pretrained models in the near future.

Best, A

alexklwong avatar Sep 28 '23 18:09 alexklwong

Hi, thank you for the suggestion about kbnet. I will try on it in these days. I believe your work will greatly support my research

myboyhood avatar Oct 01 '23 01:10 myboyhood