PointNet

PointNet is a deep neural network designed to do classification and segmentation on 3D point clouds. The network has been implemented in PyTorch.

Training the classification network

The classification network was trained for 10 epochs, both with feature transform as well as without. The weights were saved for both cases.

python show_cls.py --model ./cls_FT/cls_model_9.pth --feature_transform

Average Train Accuracy: 0.960422 Average Test Accuracy: 0.948385

python show_cls.py --model ./cls_noFT/cls_model_9.pth

Average Train Accuracy: 0.976830 Average Test Accuracy: 0.957163

Training the segmentation network

The segmentation network was trained for 10 epochs, both with feature transform as well as without. The weights were saved for both cases. This was done for two classes: Chair and Guitar.

‘Chair’ Class:

python train_segmentation.py --dataset ./shapenet/ --nepoch 10 --class_choice Chair --feature_transform

Testing mIOU for class Chair: 0.8156666563399274

Training mIOU for class Chair: 0.7940363590016127

python train_segmentation.py --dataset ./shapenet/ --nepoch 10 --class_choice Chair

Testing mIOU for class Chair: 0.8010046474006883

Training mIOU for class Chair: 0.7939503764027812

‘Guitar’ Class:

python train_segmentation.py --dataset ./shapenet/ --nepoch 10
--class_choice Guitar --feature_transform

Testing mIOU for class Guitar: 0.7910344983065059

Training mIOU for class Guitar: 0.8321464373591417

python train_segmentation.py --dataset ./shapenet/ --nepoch 10
--class_choice Guitar

Testing mIOU for class Guitar: 0.8199666335989975

Training mIOU for class Guitar: 0.8280420213170521

Segmentation Outputs for 'Chair' and 'Guitar' classes