Bug related to ProjectInseg about scenegraphfusion

[SgtVincent]

Update:
If flag --rendered is set to 0, the loaded depthImage has a shape of (224, 172) while colorImage has a shape of (540, 960). If flag --rendered is set to 1, depthImage will be reshaped to (540, 960) by the reindeer before being sent to inseg_->ProcessFrame() method. I wonder if this mismatch of sizes could lead to segmentation fault.

[ShunChengWu]

Hi,
The image sizes are correct for 3RScan dataset. We currently only support with rendered view.
If you want to use the original depth image you will need to resize it to be the same size as the RGB image. But we don't encourage that since the depth maps from 3RScan dataset are kind of bad.

[SgtVincent]

Thanks for your answer! Yet it really confused me that you assert dataset == 'ScanNet' in RUN_GenSeg.py script. I thought you would ONLY use non-rendered mode for 3RScan dataset. See https://github.com/ShunChengWu/3DSSG/blob/24d4ac520c3441fb6815d0d4890dd845e10e5cc7/script/RUN_GenSeg.py#L113

[ShunChengWu]

previously the program didn't support the online generation of the rendered view of 3RScan. So you would need to generate the rendered view first then run our code. We added this feature later but forgot to change that part in the script. the rendered view should always be true unless you want to train with your own settings. I will change that soon.
Thanks

[SgtVincent]

Problem solved, thanks!