In this RCNN Object Detection notebook, RCNN is used to detect object. The object to be detected is raccoons.
200 raccoon images and their corresponding annotations (ground truth bounding box) are used to train the RCNN.
The steps to perform object detection with RCNN:
- Create image dataset consisting of 2 classes; each of which contains images of raccoon and no raccoon
- Train a classifier (CNN) to classify raccoon and no raccoon
- Detect raccoon
The object to be detected here is raccoons.
The notebook is run with Tensorflow 2.6.0 and OpenCV 4.1.2
Selective search is used to create (propose) regions (an area) from an image.
Each region is compared to the ground truth bounding box, specifically by calculating the intersection of union (IOU) between the region proposed to the ground truth bounding box of a raccoon. For regions with IOU > 0.7, they are collected to class raccoon, whereas the regions with IOU < 0.05 are collected to class no raccoon.
 |
|---|
The regions proposed by the selective search |
 |
|---|
The raccoon and no raccoon dataset |
A CNN is trained to classify 2 classes; raccoon and no raccoon.
A pretrained MobileNetV2 is used with all the weights frozen. On top of it, a bunch of dense layers are added.
With only 5 epochs, the loss value for both train and validation set is already so low and the accuracy values are reaching 99%.
 |
|---|
| Learning curve of the CNN classifier |
The selective search algorithm is once again used to propose region from the image to be detected, creating region of interests (roi) which are fed as inputs to the classifier. Afterwards, the classifier classifies if the regions belong to raccoon or no raccoon. For a region that is classified as raccoon, a bounding box is drawn. At the end, non maximum suppression algorithm selects one best bounding box out of many overlapping boxes drawn.
 |
|---|
| The detection: before (left) and after non maximum suppression(right) |
The code and raccoon images used in this notebook are from the pyimagesearch tutorial.