Answers are grounded in this paper and cite their sources.
Background Detecting and localising long and thin objects has historically been a computationally intensive task. These objects often lack distinct features, occupy minimal space in images, and can be challenging to differentiate from the background. While computer vision algorithms such as You Only Look Once (YOLO) have significantly improved object detection accuracy, they still require extensive datasets and considerable training to effectively identify small or thin objects. In this article, we introduce a novel method for detecting long and thin objects, optimising both dataset creation and detection algorithms. Inspired by the YOLO model, we implement key modifications to enhance accuracy and performance, addressing its limitations in detecting such objects. The proposed method has diverse real-world applications, including identifying cracks in rock formations, guiding bridge construction, and improving missile targeting systems. For instance, an aeroplane fuselage serves as the elongated object whose precise localisation the model aims to achieve. Additionally, this technique can be utilised for any challenges related to detecting and localising elongated objects, making it a powerful tool in computer vision. Methods We employ a lightweight convolutional network for our approach. The dataset annotation is influenced by YOLO techniques, but all grid cells occupied by the object are annotated with the corresponding class number and position. We introduce a new loss function aimed at penalising localisation errors. Furthermore, we propose a novel method for calculating mean average precision (mAP) for point-based detection systems, like Long Object Correlation-based Analysis and Localisation System (LOCALS), which will be compared to existing YOLO models that utilise mAP as a performance metric. Results LOCALS achieves a mAP of 0.93 over distance thresholds of [0.5–0.95]. In comparison, YOLOv8 reaches an mAP of 0.80 over IoU thresholds of [0.5–0.95]. This suggests that our method offers significant improvements in both classification and object detection capabilities. To better understand the applicable scenarios and failure modes of LOCALS, another dataset (the rivers dataset) was constructed from 800 aerial images of rivers obtained from Kaggle, resulting in an mAP of 0.7849.
No takes yet. Share an insight, caveat, or question.
Arumugam et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: