Closed pushkalkatara closed 3 months ago
It'd certainly be useful. While don't have this in our current milestone, we'd be happy to merge if someone made a PR.
i would like to work on the issue @krrish94 , I think we can integrate the S3DIS
dataset rather than the full 2D-3D-S dataset.
Further when the DGCNN Semantic Segmentation model is added, we can create an example using the S3DIS dataset. Let me know your thoughts on this.
That sounds good to me. You can proceed to work on this issue
Hi @krrish94 , please find minimal dataset loading here - https://github.com/pushkalkatara/kaolin/blob/s3dis/kaolin/datasets/s3dis.py
I had few doubts: The structure of the dataset is as follows: Areas -> Rooms -> ObjectName.txt Annotations -> XYZRGB. The annotations has XYZRGB format, in which XYZ is the pointcloud, RGB is the image. In the current implementation (focused to develop dataset for semantic segmentation), given a room id, it would load all objects in the room and return labels (object names) and point clouds. This is feeded as an input to a sem-seg model.
As it would be better to implement a generalized dataset rather than a specific task, it would be helpful if you could suggest the design pattern which I should follow for loading the dataset.
Currently 2D-3D-S dataset is not part of kaolin.datasets. Would it be a useful addition?
Hi! Unfortunately this link is not available anymore. Could you please let me know how I can download the S3DIS dataset?
Closing old issue.
Currently 2D-3D-S dataset is not part of kaolin.datasets. Would it be a useful addition?
http://buildingparser.stanford.edu/dataset.html#Download