The recognition of particular objects instances (e.g. my coffee cup or my wallet) is an important research topic in robotics, as it enables tasks like object manipulation in domestic environments in real-time. However, in recent years most efforts have been aimed to solve generic object detection and object class recognition problems. In this work, a method for performing recognition of particular objects instances, named YoloSPoC, is proposed. It is based on generation of high-quality object proposals by using YOLOv3, computing descriptors of these proposals using a MAC (Maximal Activation of Convolutions) based approach, recognizing the object instances using an open-set nearest neighbor classifier, and filtering of overlapping recognitions. The proposed method is compared to state-of the-art methods based on local features (SIFT and ORB based methods) using two datasets of home-like objects. The obtained results show that the proposed method outperforms existing methods in the reported experiments, being robust against conditions like (i) occlusions, (ii) illumination changes, (iii) cluttered backgrounds, (iv) presence of multiple objects in the scene, (v) presence of textured and non-textured objects, and (vi) object classes not available when training the proposal generator.
CITATION STYLE
Loncomilla, P., & Ruiz-del-Solar, J. (2019). YoloSPoC: Recognition of Multiple Object Instances by Using Yolo-Based Proposals and Deep SPoC-Based Descriptors. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 11531 LNAI, pp. 154–165). Springer. https://doi.org/10.1007/978-3-030-35699-6_12
Mendeley helps you to discover research relevant for your work.