SPD: Semi-Supervised Learning and Progressive Distillation for 3-D Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• 3-D object detection accuracy is limited by annotation quality and dataset sparsity.
• Existing methods struggle with high detection accuracy across diverse scenarios and classes.

Purpose of the Study:

• To propose a novel method, semi-supervised learning and progressive distillation (SPD), to improve label efficiency in 3-D object detection.
• To enhance generalization and mitigate overfitting using specific augmentation and loss functions during training.

Main Methods:

• SPD utilizes two large backbones with periodic IO augmentation (PA) for both labeled and unlabeled data.
• A data scale and ratio-sensitive loss (data-loss) is employed to balance large unlabeled datasets and limited labeled data.
• Progressive distillation (PD) compresses the trained teacher backbone into a smaller student model, preventing performance degradation.

Main Results:

• SPD achieves 0.32 higher accuracy than fully supervised VoteNet using only 50% labeled data and a 27% smaller model.
• With only 2% labeled data, SPD achieves comparable accuracy to fully supervised PV-RCNN while reducing inference time by 30%.

Conclusions:

• SPD significantly improves label efficiency and model performance in 3-D object detection.
• The method demonstrates effectiveness across various indoor and outdoor datasets, offering a practical solution for data-scarce scenarios.