jamycheung / 360bev Goto Github PK

Hello,

Firstly, I'd like to thank you for sharing this interesting codebase. I have been reviewing the BEV360_segnext_s2d3d.py script and have a couple of questions I hope you could clarify for me:

Could you please explain how the masks_inliers variable is derived in the code? I'd like to understand the logic behind it.

Regarding the variable m, it seems to be derived in a rather unusual way. The threshold_index_m is set as the maximum value of proj_index. As a result, the boolean mask m will mark all elements in proj_index that are smaller than this maximum value as True. The only exception, where m would be False, is when an element in proj_index equals threshold_index_m. Could you please elaborate on the reasoning behind this approach?

I appreciate your time and assistance in advance. Looking forward to your response.

Best regards

Dear Authors:
When I downloaded the 360-FV dataset, I had found it didn't include label file and it had 42 numbers in .png. Besides, The dataset includes "self.df = pd.read_csv("eigen13_mapping_from_mpcat40.csv")", it is hard to find the .csv for me. Would you mind sending it to me？

Bests.

Dear authors,

thanks for your excellent work! I have some questions about the 360FV-Matterport dataset. It would be great if you can help me out.

1. The original Matterport dataset has 90 scenes in total, but I can only find train61+val7+test18=total86 scenes in 360BEV paper. Where are the rest 4 scenes? Do I miss something?
2. I cannot figure out the train/val/test subset division of the 360FV-Matterport dataset. Would you mind giving me some suggestions on the subset division?

Thank you in advance for your help.

Best,
Junwei

I noticed that the project makes use of a checkpoint file named mit_b2.pth. Could you please provide some information about its origin? Specifically:

Where was this checkpoint file downloaded from?
Is it a pre-trained model that does not require further training to be used?
Thank you for your time and effort on this project. Looking forward to your response.

Hello,
The download from Google Drive doesn't contain this folder: 360BEV-Matterport/valid/smnet_training_data_zteng.

Is this a folder I need to create myself?
Looking forward to your reply! Thanks!

Your work is very good, but I don't know how to organize the data you provide in Google Drive into a suitable format and test it

In the encoder section of your code, I noticed that the index doesn't appear to be utilized. Instead, you seem to recompute the corresponding panorama index based on the mask height. Given that you've already calculated smnet_training_data_maxHIndices, why isn't this index used directly instead of recalculating it?

Dear authors,
Hello, I greatly admire your impactful and innovative work: 360BEV：Panoramic Semantic Mapping for Indoor Bird’s-Eye View
When I read your excellent work, I saw:
"The original Matterport3D [5] was collected via narrow-FoV cameras. As shown in Fig. 3, we convert the 18 narrow-view images and annotations into the 360° format by using rotation-translation matrices."
Could you please provide the process code of acquiring Matterport3D panoramic segmentation image?
Thanks！