Show HN: I achieved over 10% improvement on 3D PointCLIP

achieved more than 96 accuracy on modelnet40 test dataset with lightweighted custome model pipeline. Projecting (added gaussian splatting) 3d pointcloud dataset into 2d image and putting them to clip vit-b/16

1. python==3.10
2. git clone this repo
3. cd 3d-vlm-gaussian-splatting-classification-on-modelnet40
4. pip install -r requirements.txt
5. python -m inference.py
6. if it's first time running this model, then in config.py, set exist_flag = True
7. if you want to do inference on more dataset, you can set max_samples in config.py

1. torch in requirements.txt is cpu only. change it if you want to use cuda
2. if you want to use cuda version, delete torch in requirements.txt and do (3)
3. pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu<version you want e.g) 118 if cuda 11.8>
4. sample test accuracy is 70% and overall validation performance for test datasrt is 95 for this pth
5. you can run this even in only cpu env

Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang and J. Xiao 3D ShapeNets: A Deep Representation for Volumetric Shapes Proceedings of 28th IEEE Conference on Computer Vision and Pattern Recognition (CVPR2015) Oral Presentation · 3D Deep Learning Project Webpage