Related Experiment Video
Updated: Sep 8, 2025
06:54
Photorealistic Learned Landscapes for Augmented Reality
Published on: June 27, 2025
155
SPOT: Scalable 3D Pre-Training via Occupancy Prediction for Learning Transferable 3D Representations.
Summary
Scalable Pre-training via Occupancy prediction (SPOT) enables efficient 3D representation learning for LiDAR point clouds. This approach reduces labeling burden and improves performance across diverse tasks and domains.
Area of Science:
- Computer Vision
- Machine Learning
- Robotics
Background:
- Annotating 3D LiDAR point clouds for perception is crucial but labor-intensive.
- Pretraining-finetuning methods can reduce labeling efforts for downstream tasks.
Purpose of the Study:
- Introduce SPOT (Scalable Pre-training via Occupancy prediction) for learning transferable 3D representations.
- Demonstrate the effectiveness of SPOT in a label-efficient fine-tuning paradigm.
Main Methods:
- Developed SPOT, a pre-training strategy using occupancy prediction for transferable 3D representations.
- Implemented beam re-sampling for point cloud augmentation and a class-balancing strategy to address domain gaps.
- Investigated scalable pre-training with increasing amounts of pre-training data.
Main Results:
- SPOT shows effectiveness across various public datasets and downstream tasks.
- Demonstrated general representation power, cross-domain robustness, and data scalability.
- Confirmed that more pre-training data leads to better downstream performance.
- Showcased compatibility with unlabeled data.
Conclusions:
- Occupancy prediction is a viable task for general representation learning in 3D point clouds.
- SPOT facilitates understanding of LiDAR data and advances LiDAR pre-training.
- The proposed methods address domain gaps and enhance the scalability of pre-training strategies.
