SNIC SUPR
FasterRCNN for Stump Detection
Dnr:

SNIC 2019/5-18

Type:

SNIC Small Compute

Principal Investigator:

Ahmad Ostovar

Affiliation:

Umeå universitet

Start Date:

2019-02-07

End Date:

2020-03-01

Primary Classification:

10201: Computer Sciences

Webpage:

Allocation

Abstract

Stumps Detection using Faster R-CNN: To create a convolutional neural network (CNN) the input, middle, and final layers are stacked on top of each other. The input layer has the size of [32 32 3]. The middle layer includes convolutional, rectified linear units (Relu), pooling layers with filter size [3 3], and a total of 32 filters. The final layer contains blocks of fully connected layer, Relu layer, softmax layer, and classification layer. configuring the training options and training the Faster R-CNN detector. To train the Faster R-CNN detector we used the ‘trainFasterRcnnObjectDetector’ function from Computer Vision Toolbox in MATLAB. It trains the detector in four steps. The first two trains the region proposal and detection networks and the final two steps combines the trained networks (from the first two steps) to create a single network for detection. Therefore, each training step can have different training configurations. In this work we set all the training options equally for all the steps, except the learning rate which was set to a lower value in the final two steps, because they are fine-tuning steps and the network weights can be modified more slowly. (We use Matlab with Deep Learning Toolbox)