Abstract

Object detectors based on deep neuralnetworks have revolutionized the way we look forobjects in an image, outperforming traditional im-age processing techniques. These detectors are of-ten trained on huge datasets of labelled images andare used to detect objects of different classes. We ex-plore how they perform at detecting custom objectsand show how shape and deformability of an objectaffect the detection performance. We propose an au-tomated method for synthesizing the training imagesand target the real-time scenario using YOLOv3 asthe baseline for object detection. We show that rigidobjects have a high chance of being detected withan AP (average precision) of 87.38%. Slightly de-formable objects like scissors and headphones showa drop in detection performance with precision aver-aging at 49.54%. Highly deformable objects like achain or earphones show an even further drop in APto 26.58%.

Reference

Djukic, N., Kropatsch, W., & Vincze, M. (2020). The Difficulties of Detecting Deformable Objects Using Deep Neural Networks. In Proceedings of the Joint Austrian Computer Vision and Robotics Workshop 2020 (p. 6). https://doi.org/10.3217/978-3-85125-752-6-30