Active learning method and device for semantic clustering of data points based on spatial structure diversity

Abstract

The present invention discloses an active learning method and device for semantic clustering of data points based on spatial structural diversity. The method comprises the following steps: 1) grouping the original point cloud into individual superpoints, randomly selecting a small number of superpoints for manual labeling, and training a preliminary model; 2) using a weighted superpoint uncertainty estimation method and a spatial-structural diversity inference method to select candidate superpoints to be manually labeled; 3) utilizing a noise-aware iterative labeling strategy to segment candidate superpoints with purity below a threshold into multiple subregions for processing. The candidate superpoints are manually labeled and the model is fine-tuned; 4) calculating three metrics, softmax entropy, structural complexity, and color discontinuity, for the remaining unlabeled superpoints, and assigning pseudo-labels to a group of unlabeled superpoints with high regional information scores, which become labeled data; 5) returning to step 2, retraining or fine-tuning the model using the labeled data, and repeating the cycle until the manual labeling budget is exhausted.

Images