Is it possible to train SoftGroup on our own dataset?

[bilel-bj]

I have a dataset a point cloud. I labeled it using CloudCompare and add two Scalar Field values: the first is the semantic class and the second is the instance class.
Could I train it on my dataset?

[thangvubk]

Yes. As long as your data has semantic_id and instance_id you can use this repo to train your data. To do that you need to create new data directory in SoftGroup/dataset/ then preprocess to generate .pth file for input and .txt file for ground truth (example here and here).

After that you can create a new dataloader in SoftGroup/data, follow scannetv2_inst.py.

[bilel-bj]

Thanks a lot for your response. Do you have a specific tool for generating instance segmentation labels of the dataset? Or should I keep using Cloud Compare by adding two scalar fields to every point: semantic_id and instance_id

[thangvubk]

I just used standard public dataset. I think you can use that tool to label your data.

[bilel-bj]

Thanks a lot for your reponse. Do you think that this format is valid for training. This is a sample of the txt file

-------------------------------------------------------------------------------------------------
X   Y   Z    RGB semantic_id instance_id
-------------------------------------------------------------------------------------------------
-87.30151367 48.11581421 -32.99898148 0 0 0 1.000000 0.000000
-83.23958588 53.01259232 -31.98804665 0 0 0 1.000000 0.000000
-91.00957489 52.23581314 -25.49759102 0 0 0 1.000000 0.000000
-91.01223755 52.74370193 -25.20007706 0 0 0 1.000000 0.000000
-82.53973389 53.23898315 -32.66012573 0 0 0 1.000000 0.000000
-82.68397522 51.79740906 -34.88851547 0 0 0 1.000000 0.000000
-82.80941010 55.73773575 -32.12082672 0 0 0 1.000000 0.000000
-83.50354767 54.06523132 -31.11934280 0 0 0 1.000000 0.000000
-83.09005737 53.63531876 -31.71231270 0 0 0 1.000000 0.000000
-83.52418518 56.01091003 -30.02872658 0 0 0 1.000000 0.000000
-92.61465454 47.71725464 -30.11337662 0 0 0 1.000000 0.000000
-85.50755310 50.83419800 -31.79884911 0 0 0 1.000000 0.000000
-85.05304718 49.18477631 -34.39563751 0 0 0 1.000000 0.000000
-87.02261353 47.79521561 -33.57074738 0 0 0 1.000000 0.000000
-87.79747009 48.99874496 -31.63723564 0 0 0 1.000000 0.000000
-88.17048645 48.41600037 -32.07061386 0 0 0 1.000000 0.000000
-83.23866272 51.24077225 -34.98325729 0 0 0 1.000000 0.000000
-88.18229675 47.79760361 -32.80801010 0 0 0 1.000000 0.000000
-91.42159271 48.30610657 -30.03435135 0 0 0 1.000000 0.000000
-87.46870422 47.66637421 -33.31636047 0 0 0 1.000000 0.000000

[thangvubk]

Yes. This is expected format.

[gfiameni]

On the same topic, I have an annotated 2D (RGB-D) COCO datasets, is there any way I can leverage SoftGroup for 3D segmentation? Tnx

[thangvubk]

@gfiameni First, you need to convert RGBD to point clouds format. You may check here as reference.