Towards novel reliable landmarks in robotics: evaluation of graph-based methods for rectangle detection in images
Rectangular elements are very common features in structured man-made environments and they provide high-level distinctive information, very useful for object and place recognition tasks. In this article, we propose and evaluate several rectangle detection methods based on the Rectangle Intersection-based Detector using Graphs and Elongation (RIDGE) Vega-Martínez et al. (2024); de la Puente et al. (2025). The novel versions proposed of the RIDGE algorithm are a) RIDGE-Airline: a version that replaces the underlying line detector by a recent learning-based method, b) RIDGE-ROI: a version that extracts Regions of Interest (ROIs) by classical methods and then applies RIDGE with redundancy checks, c) RIDGE-YOLO: a version that first applies semantic bounding box detection based on YOLO and subsequently extracts rectangle vertices by means of the original RIDGE method. As a complementary exploratory variant, we additionally introduce and evaluate d) YOLO-BR (Boundary Refinement), which constructs quadrilateral candidates using a directed graph of supporting lines and selects boundaries through finite-segment evidence and proposal association. For qualitative and quantitative evaluation of the different methods, we have created and released two datasets with rectangle annotations: one based on synthetic realistic simulation environments and the other based on a selection of images from a publicly available robotics dataset.
Authors
- Paloma de la Puente (ORCID: https://orcid.org/0000-0002-8652-0300)
- Jiaxi Huo
- Marco David Arias
- Zhenhao Zhang
Institutions
- Politecnico di Torino (IT)
- Centre for Automation and Robotics (ES)
- Universidad Politécnica de Madrid (ES)
Publication Details
- Journal
- International Journal of Intelligent Robotics and Applications
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1007/s41315-026-00610-2
- Primary Topic
- Image and Object Detection Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00