Point-MPP: Point Cloud Self-Supervised Learning From Masked Position Prediction.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Masked autoencoding improves fine-tuning but often focuses on low-level details.
• This leads to weak transfer capability due to a lack of high-level semantics.

Purpose of the Study:

• To develop a novel pretraining method for point clouds that captures high-level semantic information.
• To enhance the transfer capability of models for downstream vision tasks.

Main Methods:

• A jigsaw puzzle-inspired approach that masks point cloud patch positions instead of content.
• Utilizing a Transformer-based model to regress the positions of masked patches.
• Implementing a consistency constraint on latent representations to enforce semantic cues.

Main Results:

• The proposed pretraining scheme enables Transformer models to capture discriminative point cloud semantic information.
• Outperforms previous state-of-the-art methods across six downstream vision tasks.
• Demonstrates superior performance in learning transfer-robust semantics.

Conclusions:

• Predicting masked patch positions is an effective strategy for learning high-level semantics in point clouds.
• The method significantly enhances performance on various downstream vision tasks.
• Achieves new state-of-the-art results, highlighting the potential for improved point cloud understanding.