Self-Supervised 3D Shape Representation Learning with PyTorch and ShapeNet Dataset
This project focuses on self-supervised representation learning of 3D shapes using point clouds from a subset of the ShapeNet dataset. Implemented with PyTorch and PyTorch Geometric (PyG), it employs the DynamicEdgeConv approach with a reduced architecture of two layers. The model is trained using InfoNCE/NT-Xent loss from the PyTorch Metric Learning library, with a temperature parameter for similarity measure balancing. Visualizations are provided using Plotly, demonstrating successful contrastive learning and embedding space arrangement.
Read the blog on this project with detailed description : https://medium.com/@singhamit_/self-supervised-contrastive-learning-for-shape-representation-learning-with-pytorch-and-shapenet-7db494826bf2