Regional Proposal Network

Varied Technologies Used in Object Detection

Varied technologies have been used to build object detection machines that have gained popularity in face recognition, video surveillance, tracking the movement of a ball during football matches, video object co _x0096_ segmentation, face detection, and other computer vision tasks. Current technologies such as Region Convolution Neural Network (R - CNN) and SPPnet which are using detection network have been adopted to replace the state -of - the - art object detection machines built on Support Vector Machine Technology. This paper presents a report on the implementation of Region Proposal Network. The device can share full _x0096_ image convolution features with the detection networks (Zhaowei and Vasconcelos 2018).

Introduction

With the advancement in technology, object detection is optimized with First Regional Convolution Neural Network (FR- CNN) and Regional Proposal Network (RPN). The earlier Regional Convolution Neural Network was expensive concerning time for image processing hence an improvised model of this was First Regional Convolution Neural Network (FN - CNN) which achieves near real-time image processing using deep networks. This paper discusses how Firs Region Convolution Networks shares full image convolution features with the Regional Proposal Networks as a way of attaining real time image processing for object detection (Aleksis and Sminchisescu 2018).

Region Proposal Network and Selective Search

Region Proposal Network uses Selective Search methods which rely on inexpensive feature with economic interference scheme. The technique merges image pixels by utilizing low _x0096_ level engineering features. However, the comparison between Selective Search and some efficient detective networks shows that Selective Search is slower by 2 seconds when executing an object image in the Central \u00a0Processing Unit. An improvement for Selective Search method was achieved by adopting EdgeBox methods that execute object image at the rate of 0.2 seconds per picture. Even though Regional Proposal Networks completes nearly real-time object detection, training on the technique shows that it consumes much time when using detection networks.

Design and Training of the Regional Proposal Network

According to ( Bharat and \u00a0Davis 2018) the Regional Proposal Network was designed to take an image of any given size as an input and produce a set of rectangular object proposals as output with each rectangle having its abjectness score. The Regional Proposal Network was a model to achieve real-time object detection by using convolution networks. During training of the Regional Proposal Network, Zeiler and Fergus’s models containing five sharable convolution layers and Simonyan and Zisserman model containing thirteen sharable convolution networks were tested to simulate the detection process. During the production of Regional Proposal Network a small system was slid over the convolution feature map output of the last sharable convolution layer.

Methodology

During training and testing of the Regional Proposal Network, PASCAL VOC 2007 detection benchmark was used. The dataset consisting of about 5 kilo trainval images and 5-kilo test images from different object categories were fed into the PASCAL VOC 2007 detection benchmark, and the results were noted. For performance comparison purposes, PASCAL VOC 2012 benchmark was used to rescale the output of the PASCAL VOC 2007 benchmark.

Ablation Experiments and Conclusion

Several ablation experiments were conducted to investigate the behavior of the Regional Proposal Networks. The first ablation experiment was done to show effectiveness in sharing convolution layers between the Regional Proposal Convolution Network RPN and the Fast Regional Convolution Neural Network (FR - CNN). The second ablation experiment was done to check on the performance of the Regional Proposal Network by disentangling the influence of Regional Proposal Network on training the Fast Regional Convolution Neural Network. This was achieved by training the Fast Regional Convolution Neural Network using the 2 kilos Selective Search (SS) proposal and ZF net. Then the detector was fixed, and the evaluation of the detection mAP was done by changing the proposal region used in testing the time. The experiment showed that when the Selective search replaced with 300 Regional Proposal Network proposal during test time, 56.8% of the mAP was achieved and the loss in this mAP was brought about by the incontinence between the training data set or test proposals.

Conclusion

There has been tremendous increase in technology used in developing object detection computer-based programs. The older technologies such as Regional Convolution Neural Network lead to the development of time-consuming object detection machines. The introduction of Fast Regional Convolution Neural Network that shares convolution layers with the Regional Proposal Networks has resulted in the development of faster, efficient, and convenient object detection machines that bring about real-time image detection and processing. Object detection technology is being applied in face detection, face recognition, video surveillance, tracking the movement of a ball during football matches, video object co _x0096_ segmentation, face recognition, and other computer vision tasks (Wu and Hoi 2018).

Work Cited

Pirinen, Aleksis, and Cristian Sminchisescu. “Deep Reinforcement Learning of Region Proposal Networks for Object Detection.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.

Cai, Zhaowei, and Nuno Vasconcelos. “Cascade R-CNN: delving into high quality object detection.” IEEE CVPR. 2018.

Singh, Bharat, and Larry S. Davis. “An Analysis of Scale Invariance in Object Detection_x0096_SNIP.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.

Sun, Xudong, Pengcheng Wu, and Steven CH Hoi. “Face detection using deep learning: An improved faster RCNN approach.” Neurocomputing 299 (2018): 42-50.

.