Less is More: Reinforcement Learning-powered Training Data Quality Evaluation and Sampling for Enhancing Computer Vision in Construction

초록

Deep learning-based computer vision in construction has often assumed that more training data leads to better performance. This quantity-driven approach overlooks the detrimental effects of redundant, noisy, or mislabeled samples that can degrade model accuracy while inflating computational costs. To overcome this limitation, we propose and test an alternative hypothesis: selectively removing low-quality training samples can outperform indiscriminate use of all data. Specifically, we present a reinforcement learning (RL)-powered data quality evaluation framework that scores individual training images and prioritizes the most beneficial samples for model learning. Within a unified environment, the RL agent (referred to as the RL-evaluator) interacts with an object detector by probabilistically selecting subsets of high-quality samples for model training and using validation performance as a reward signal to optimize its selection policy. Experiments on a construction benchmark dataset of varying sizes show that removing low-quality samples identified by the RL-evaluator consistently improves detection performance compared to training on the full dataset, while also reducing computational overhead. These findings demonstrate the proposed method as a scalable and robust way for enhancing computer vision in construction.

키워드