Exploring Vision-Based Active 3D Object Detection by Informativeness Characterization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Vision-based 3D object detection (3DOD) is cost-effective but requires extensive data annotation.
• Active learning (AL) can minimize annotation efforts by prioritizing informative samples.
• Exploring AL for vision-based 3DOD is crucial for efficient model training.

Purpose of the Study:

• To propose a novel active learning framework, HMAD, for vision-based 3DOD.
• To address the challenge of labor-intensive data annotation in 3DOD.
• To maximize model performance with limited labeled data.

Main Methods:

• Developed HMAD framework incorporating Height Modeling and Adaptive Diversity-based sampling.
• Introduced a height-guided adversarial module in Bird's-Eye View (BEV) space for 2D-to-3D mapping informativeness.
• Proposed Budget-aware SpatioTemporal diversity Sampling (BSTS) and Class Balance Sampling (CBS) for comprehensive sample characterization.

Main Results:

• HMAD achieves comparable performance to traditional methods using only 50% of annotated training data.
• The framework effectively characterizes sample informativeness across input, feature, and prediction spaces.
• Demonstrated strong generalization capabilities across different conditions.

Conclusions:

• HMAD offers an efficient solution for reducing annotation costs in vision-based 3DOD.
• The proposed sampling strategy effectively identifies and selects the most informative samples.
• This approach significantly enhances the practicality of deploying 3DOD systems.