Cannot find the classification engines in the master branch

[praveenv4k]

Hi, I noticed that many 3D deep learning architectures were implemented until the reformatting of kaolin is done. I cannot find them in the latest master branch. Will they added eventually as part of this repo or will they moved to another repo?

[Caenorst]

Hi @praveenv4k ,

As stated in the change logs model zoo will part of another repo.

[MBrandt-NASA]

Is there any timeline for when the Model Zoo and 3D DL "task" examples / tutorials (classification, segmentation, object detection, etc.) will be available?

I wanted to use Kaolin in my projects, but I'll have to move to another 3D DL library for the time being, until your examples and Model Zoo are released for the 0.9 version of Kaolin.

[Caenorst]

@MBrandt-NASA Do you have some model in mind? It would help us prioritize.

[GSFC-587]

Hi @Caenorst thanks for asking :).

This year I'm trying to accomplish 3 tasks with 3D DL:

1) 3D object detection w/ bounding boxes. (e.g. detecting a crater in lunar terrain), or potentially semantic segmentation, but I prefer objects w/ bounding boxes 2) pose estimation (e.g. estimating the pose of a known model -- satellite in orbit) 3) map generation / registration -- matching an incoming terrain patch (from a Lidar scan) with the larger containing terrain (e.g. a lunar rover, w/ Lidar, traverses a large terrain and generates a terrain map in real time as each Lidar scan comes in)

I'm evaluating models for: simplicity, reliability, and performance (speed and accuracy)

---------------- Potential models ----------------

Object Detection:

• PointPillars / SECOND (https://github.com/nutonomy/second.pytorch, https://pdfs.semanticscholar.org/5125/a16039cabc6320c908a4764f32596e018ad3.pdf)
• YOLOv4 (https://github.com/maudzung/Complex-YOLOv4-Pytorch)
• VoteNet (https://github.com/facebookresearch/votenet)
• PVR-CNN (https://arxiv.org/abs/1912.13192)
• PVCNN (NVIDIA Jetson spotlight -- https://news.developer.nvidia.com/point-voxel-cnn-3d/?ncid=so-twit-99540#cid=em02_so-twit_en-us, https://github.com/mit-han-lab/pvcnn)

Pose Estimation:

• PCRNet -- uses PointNet (https://github.com/vinits5/pcrnet_pytorch)

Registration:

• PointNetLK -- uses PointNet (https://github.com/hmgoforth/PointNetLK)
• DeepMapping (https://github.com/ai4ce/DeepMapping)

I'm open to other models if you all know others that are simpler or have higher performance, for the above 3 tasks.

Btw, I'm trying to prototype on Windows 10 and potentially deploy to Linux if needed, so it would be nice if both OS's are supported, as with Kaolin 0.1.

I really enjoyed working with Kaolin 0.1 -- PointNet and PointNet++ both worked great for me! So I'm looking forward to seeing task examples / tutorials for Kaolin 0.9.

Very important -- examples / tutorials on how to adapt custom datasets to Kaolin 0.9.