yuxumin/PoinTr

What it solves

PoinTr addresses the problem of point cloud completion, where a model must predict the missing parts of a 3D object from an incomplete or partial point cloud. It specifically aims to handle diverse and challenging real-world scenarios where point clouds may be significantly incomplete.

How it works

The project uses a transformer-based architecture. It represents a point cloud as a set of unordered groups of points with position embeddings. These are converted into a sequence of point proxies, which are then processed through a transformer encoder-decoder architecture to generate the completed 3D shape.

Who it’s for

This tool is designed for researchers and developers working in 3D computer vision, robotics, and geometric deep learning who need to reconstruct full 3D objects from partial sensor data.

Highlights

• Geometry-Aware Transformers: Utilizes a transformer encoder-decoder to generate completed point clouds.
• Diverse Benchmarks: Introduces ShapeNet-55 and ShapeNet-34 datasets to better reflect real-world incomplete point cloud scenarios.
• AdaPoinTr: Includes an advanced version featuring Adaptive Denoising Queries for state-of-the-art performance.
• Baseline Implementations: Provides PyTorch implementations of other baseline models like GRNet, PCN, TopNet, and FoldingNet for comparison.