A Jittor implementation of Point Cloud Transformer (PCT) for ModelNet40 point cloud classification.
Introduction
This project is for Computer Graphics PA3. The goal is to train a point cloud classification model on the ModelNet40 dataset using the Jittor deep learning framework.
The input of the model is a 3D point cloud. Each point has three coordinates (x, y, z). The network predicts one of 40 object categories for each test sample.
Model
The main model is based on Point Cloud Transformer (PCT). The overall structure includes:
point-wise feature embedding with Conv1d
several self-attention layers for modeling global relationships between points
feature concatenation and fusion
global max pooling
fully connected classification head
The final output is a 40-dimensional logits vector for ModelNet40 classification.
Files
PCT_jittor/
├── pct.py # main training and testing code
├── README.md # project description
├── .gitignore # ignored files
└── LICENSE # license file
关于
A Jittor implementation of Point Cloud Transformer (PCT) for ModelNet40 classification.
PCT_jittor
A Jittor implementation of Point Cloud Transformer (PCT) for ModelNet40 point cloud classification.
Introduction
This project is for Computer Graphics PA3. The goal is to train a point cloud classification model on the ModelNet40 dataset using the Jittor deep learning framework.
The input of the model is a 3D point cloud. Each point has three coordinates
(x, y, z). The network predicts one of 40 object categories for each test sample.Model
The main model is based on Point Cloud Transformer (PCT). The overall structure includes:
Conv1dThe final output is a 40-dimensional logits vector for ModelNet40 classification.
Files