wz0919 / scalevln Goto Github PK

Hello @wz0919 @YicongHong @jialuli-luka , I want to express my gratitude for your outstanding contributions to the field of VLN.
Could you please release the script for extracting features for images as you have used additional environment HM3D, Gibson..
so could you please tell in detail how you have extracted features.

Thank you so much for your time.

Hey, thank you for the amazing work. Can you pls release the script to extract the below features:

• clip_vit-h14_mp3d_hm3d_gibson.hdf5
• clip_vit-b16_mp3d_hm3d_gibson.hdf5

Thank You

Hi, thanks for sharing such a great work!
I am wondering if it is possible to share the weights of the Speaker to generate navigational instructions in ScaleVLN.
I followed the paper to train an EnvDrop Speaker with the clip feature provided on the R2R dataset. However, when generating the instructions for HM3D environments with the provided feature, the results are much worse than the instructions in ScaleVLN.
Could I know what the problem is or is there any plan to release your trained speaker?
Thanks!

Hi, thanks for your interesting work!

I just wanted to ask how you installed Matterport3DSimulator. I think the command you give is for local installation without using docker (correct me if I am wrong). Is there a convenient way to use the docker version? Because I successfully installed Matterport3DSimulator with docker, but had difficulties installing the dependencies when I attempted to do it locally (because many of the packages they used are out of date).

Thanks for your time.

I am very grateful for your research on VLN. Could you please release or share the code, data, and trained models for other downstream tasks, such as REVERIE, R4R, and R2R-CE? Your work could greatly benefit my ongoing projects. Thank you so much for your time.

Thanks for this outstanding work!
Could you please share your Recovered Images from the new environments Gibson and HM3D?

Hi, @wz0919
Do you have a plan to release the depth images or the depth features?

In the "Appendices B.3.Effect of Depth Modality" of the paper, I see you compared RGB data and RGBD data.

Hi, when I upload the test file of r2r to the leaderboard, I fail to test the results, it display ' from df045272aeba414dbefac729c49d92f5 to f45a8a43423e45788bde4e50d4ec1e2e but the navigation graph contains no edge between these viewpoints ' . Did you meet this problem. Looking forward to your replay.