Imagine we want to recognize/identify an object in the images streamed from camera feeds (such as to recognize thief/suspect at the immigration checkpoint, airports, stations, etc.).
To do that, the convolutional neural network (CNN) is currently the most used method. Such popular CNN architectures such as LeNet, AlexNet, VGG, GoogLeNet, ResNet, YOLO, etc. could be retrained and then applied for inference but they are resource-intensive due to very deep network and inclusion of a huge number of parameters.
As a result, it takes substantial time to process. GPU could speed up the process. In this trial, 100 images with 100×100 pixel were sent 1-layer CNN, each to CPU and GPU. The resulted speedup is 15x.
It can be used to detect hot spots (forest firing), traffic light violation, illegal ships, traffic jam, etc. The only problem is that do we have (enough) data?