Dataset Release Issue

[tsunghan-wu]

Hi everyone,

In the recent one month, I have received a lot of emails and messages regarding to the missing datasets issue. After digging in my lab servers, PC, old laptops and so on, I eventually found a copy of "pre-processed (resize operation)" ground truth data of Matterport3D as well as ScanNetv2. After that, I contacted Yinda Zhang for the news but did not receive any reply for a while. As it can be used for academic advancement, I decided to share them although it might be a bit different from the original data.

In the following 7-14 days, I will organize these data, upload to a "safer place", and publish rules for using datasets on our GitHub. Also, feel free to contact me if you own an "original copy". I will close this issue once I finish it.

Thanks for your patience.

[tsunghan-wu]

Hi everyone,

we have released our dataset. Please follow the instructions described on our README to obtain dataset access.

Thanks.

Best,
Tsung-Han Wu

[eecoder-dyf]

Hello,
Thanks for your release! But when I open the link of the dataset https://forms.gle/aeQS12fSKPjnEuey9, google shows that I don't have the authority of this form. Could you please solve this problem?

[tsunghan-wu]

Hi @eecoder-dyf,

Sorry for the inconvenience and I have updated the permission. Thanks for your kind reminder.

Best,
Tsung-Han

[Lzshanshan]

Hello,

Thanks for your released dataset! When I compare the released resized dataset with the resized dataset generated by your code, I find they are different. And, I think the later one is better from visualized 3D point cloud. Which one should I use?

Thank you for your time.

[tsunghan-wu]

Hi @Earmu,

Sorry for the delay in my response as I've been really busy lately. For your question, I think directly using the resized data set I provided should be sufficient.

It sounds like the issue you mentioned might be caused by different resize parameters. For example, RGB images are typically best suited for bilinear interpolation, while depth images should use nearest neighbor interpolation to avoid generating interpolated depth values at the boundaries. However, this is just my personal opinion and you may want to do some additional research on this topic if you are interested.

Best,
Patrick

[Lzshanshan]

Hello @tsunghan-wu ,

Thank you for reply. Sorry for misunderstanding. Before, I meant I used the method of line 149 in the dataloader.py to resize the depth map of 1280*1024, and I found that the resized result was different from the released dataset in Google driver.

And, I think your opinions about inpterpolation method are reasonable. Now, I should modify the order parameter of skimage.transform.resize to resize depth map. (0: Nearest-neighbor, 1: Bi-linear (default)).

Thank you for your time!!!

[tsunghan-wu]

Hi @ZhenshanLiang,

Hmm... excuse me that I can't answer you why the results are inconsistent, as this work was done three years ago and I don't have the original dataset copy on hand. If you figure out why and would like to share it with me, feel free to let me know!

Hope everything goes well for your ongoing work.

Best,
Patrick

[Lzshanshan]

Hello @tsunghan-wu ,

OK! If I get some results, I will share them with you.

Thank you for your patience.

[Lzshanshan]

Hello,
Sorry for sharing late. As @tsunghan-wu said, depth images should use nearest neighbor interpolation to avoid generating interpolated depth values at the boundaries. I just want to mention that the if using skimage.transform.resize to downsample depth map, the order parameter is 0 and I use this to generate more data:
from skimage.transform import resize as numpy_resize
downsample_depth = numpy_resize(depth_map, (256, 320), mode='constant', order=0, anti_aliasing=False)