[Paper Review] ReDAL: Region-based and Diversity-aware Active Learning for Point Cloud Semantic Segmentation

Tsung-Han Wu, Yueh-Cheng Liu+5 more|arXiv (Cornell University)|Jul 25, 2021

3D Shape Modeling and Analysis30 references4 citations

TL;DR

ReDAL is a region-based and diversity-aware active learning framework for point cloud semantic segmentation that selects only informative and diverse sub-scene regions for annotation, reducing labeling costs. By combining softmax entropy, color discontinuity, and structural complexity to score regions and applying a greedy diversity-aware selection, ReDAL achieves 90% of fully supervised performance with less than 15% and 5% of annotations on S3DIS and SemanticKITTI, respectively.

ABSTRACT

Despite the success of deep learning on supervised point cloud semantic segmentation, obtaining large-scale point-by-point manual annotations is still a significant challenge. To reduce the huge annotation burden, we propose a Region-based and Diversity-aware Active Learning (ReDAL), a general framework for many deep learning approaches, aiming to automatically select only informative and diverse sub-scene regions for label acquisition. Observing that only a small portion of annotated regions are sufficient for 3D scene understanding with deep learning, we use softmax entropy, color discontinuity, and structural complexity to measure the information of sub-scene regions. A diversity-aware selection algorithm is also developed to avoid redundant annotations resulting from selecting informative but similar regions in a querying batch. Extensive experiments show that our method highly outperforms previous active learning strategies, and we achieve the performance of 90% fully supervised learning, while less than 15% and 5% annotations are required on S3DIS and SemanticKITTI datasets, respectively. Our code is publicly available at https://github.com/tsunghan-wu/ReDAL.

Motivation & Objective

• To reduce the high annotation cost of large-scale point cloud semantic segmentation by focusing labeling on only the most informative regions.
• To address the inefficiency of labeling redundant or uniform regions such as walls and floors that contribute little to model performance.
• To develop a generalizable active learning framework compatible with various deep learning architectures for point cloud segmentation.
• To mitigate label imbalance by prioritizing small, complex, or rare objects (e.g., people, chairs) over large uniform surfaces.
• To improve active learning efficiency by avoiding redundant queries of visually similar regions through a diversity-aware selection mechanism.

Proposed method

• The method uses softmax entropy to measure model uncertainty and identify regions where predictions are ambiguous.
• Color discontinuity and structural complexity are used as auxiliary signals to identify regions with high visual and geometric variation, which are more informative for learning.
• Regions are extracted from point clouds using a clustering-based method, and each region is scored using a weighted combination of entropy, color discontinuity, and structural complexity.
• A diversity-aware greedy selection algorithm penalizes similar regions in a query batch by measuring feature-level similarity in a learned embedding space.
• The framework operates iteratively: select regions, acquire labels, retrain the model, and repeat until performance saturates.
• The approach is compatible with various point cloud networks (e.g., SPVCNN, MinkowskiNet) and datasets (S3DIS, SemanticKITTI).

Experimental results

Research questions

• RQ1 Can selecting only sub-scene regions instead of full point clouds significantly reduce annotation cost without sacrificing model performance in 3D semantic segmentation?
• RQ2 How can we effectively measure the information content of a region beyond model uncertainty, using visual and structural cues?
• RQ3 To what extent does avoiding redundant queries of similar regions improve active learning efficiency in 3D point cloud settings?
• RQ4 Can a region-based and diversity-aware strategy outperform traditional scan-level active learning and random sampling in terms of label efficiency?
• RQ5 How does the proposed method handle class imbalance by prioritizing rare or complex objects during annotation selection?

Key findings

• ReDAL achieves 90% of the mean Intersection over Union (mIoU) of fully supervised learning using less than 15% of the annotations on the S3DIS dataset.
• On the SemanticKITTI dataset, ReDAL reaches comparable performance to full supervision with only 5% of the labeled points.
• The region-based selection strategy improves mIoU by over 10% compared to scan-level active learning under the same network architectures.
• The diversity-aware component prevents redundant labeling of visually similar regions, improving performance—especially under SPVCNN, where it outperforms random region selection.
• Adding color and structural complexity features improves performance on MinkowskiNet when more than 9% of points are labeled, but has minimal impact on SPVCNN.
• The method reduces label imbalance by allocating more annotation budget to small, complex objects (e.g., people, bicycles), improving their IoU by up to 29.5% under low annotation budgets.