Oriented R-CNN With Disentangled Representations for Product Packaging Detection

Abstract

Object detection is a vital task in the field of computer vision for various applications such as face detection, autonomous driving and industrial production. In recent years, with the rise of deep neural networks, there has been significant progress in improving object detection accuracy. However, despite the state-of-the-art methods being tested on public datasets, there still remains a considerable gap when applied to real-world scenarios. This is because there are many unknown types of damaged samples in industrial object detection, the scale of the types varies greatly and the position changes are complex. Many previous works focus on rotating object detection and improve it, but this paper mainly combines the prior knowledge in remote sensing and industrial scenes, and the research is more general. To fill the shortage of wrapper datasets, we established a Carton Packing Tape (CPT) Dataset with a large scale of images only containing cartons. Specifically, we first collect a large number of images of packaged cartons from the real packaging production line and provide detection boxes for them by manual labeling. We have observed that the contextual clues required for different object detection tasks exhibit inconsistency. Furthermore, targets in varying backgrounds necessitate different receptive fields, which can be dynamically adjusted using different convolutional kernels. The features naturally attended to by these receptive fields of different scales should possess a unified representation disentanglement. Based on these insights, we propose a pioneering object detection method tailored for industrial environments, termed as oriented R-CNN with disentangled representations (ORDR). The experimental results indicate that our proposed method outperforms better than some of the state-of-the-art detection techniques available.