The aim of the project is to implement some methods using the tensorflow API for image problem. The first of these was chosen to be the object_detection algorithm. The official example will first be implemented before being applied to custom datasets/classes.
The tensorflow object detection API was used to detect known items in a series of images. The API required some initial configuration as detailed in the API guidance documentation. There are several pretrained models that can be used distinguished by their trade off for peformance and speed. I used the
SSD_mobilenet_V1_coco model which was the least accurate but fastest performer to test. The two images below show the classes for ‘Person’ being idenetified with varying success.
The next step was to re-train the model used above to detect a new class in images. I chose to develop a Border Collie detector and test it using images of my dogs. I took a set of approximately 50 images from Google Images and hand labelled them with bounding boxes using LabelImg which produces xml files containing information on the image and corresponding bounding boxes.
The model was trained over a 4 hour period on a 2GB GTX750 ti Nvidia graphics card although there was little improvement after about 45 minutes. The images below show the detection in application to some pictures of Harley and Ivy.