Deep Active Learning for Efficient Training of a LiDAR 3D Object Detector

December 2019

tl;dr: Use active learning to reduce the amount of labeled data.

Overall impression

In conventional deep learning, data are labeled in a random fashion, and they are fed into a training pipeline with random shuffling.

In active learning, a model iteratively evaluates the informativeness of unlabeled data, selects the most informative samples to be labeled by human annotators and updates the training sets.

The detection task is a simplified version of classification and size/depth regression inside frustum.

Active learning is better than random baseline, regardless of the uncertainty evaluation method.

Key ideas

Technical details