Sampling Strategies for Efficient Training of Deep Learning Object Detection Algorithms

TL;DR

The paper addresses reducing labeling costs for training a deep learning object detector in laparoscopic video by introducing two sampling strategies—uniform sampling and frame-difference sampling—grounded in Lipschitz-continuity theory. It fine-tunes a pretrained YOLOv8n on data selected via these strategies and evaluates performance using IoU stability, empirical Lipschitz constants, and $\mathrm{mAP}@0.5$ on held-out data. The findings show that uniform sampling provides robust generalization at low labeling budgets, while frame-difference sampling captures dynamic frames that yield higher accuracy at larger budgets, suggesting a potential hybrid approach. This work offers a practical, end-to-end framework for low-cost domain adaptation in surgical tool detection with implications for video-based active learning in real-world imaging tasks.

Abstract

Two sampling strategies are investigated to enhance efficiency in training a deep learning object detection model. These sampling strategies are employed under the assumption of Lipschitz continuity of deep learning models. The first strategy is uniform sampling which seeks to obtain samples evenly yet randomly through the state space of the object dynamics. The second strategy of frame difference sampling is developed to explore the temporal redundancy among successive frames in a video. Experiment result indicates that these proposed sampling strategies provide a dataset that yields good training performance while requiring relatively few manually labelled samples.

Figures (4)

• Figure 1: Bean Heatmap
• Figure 2: Grasper Heatmap
• Figure 3: Testing Dataset Sample
• Figure 4: IoU stability over consecutive frames for each sampling strategy and budget. Each row shows the same label budget ($P$). Left row: Uniform Sampling; Right row: Frame Difference Sampling