I'm doing the Deep Learning Specialization, consisting of 5 courses, taught by Andrew Ng of Stanford on Coursera. Throughout the video lectures, Professor Ng often mentions and cites from papers that have been central to the fields of machine learning and deep learning.
One of these papers which, in my view, turned the field of computer vision on its head was by Redmon and colleagues (2015) in which they introduce a new approach to object detection, the YOLO algorithm. I've briefly mentioned YOLO in an earlier post.
YOLO stands for 'You Only Look Once', and as you might guess if you're not new to the field, requires only one evaluation of an image to make predictions:
"A single neural network predicts bounding boxes and class probabilities directly from full images in one evaluation." [source]
Direct end-to-end optimization is possible because the entire process of detection consists of a single neural network. If you did not read through the lines so far, I have to specify that YOLO is very good at real-time object detection. So, yes, it can be used with live video feeds. Their first basic model is able to process images at 45 fps. Moreover:
"A smaller version of the network, Fast YOLO, processes an astounding 155 frames per second while still achieving double the mAP of other real-time detectors. Compared to state-of-the-art detection systems, YOLO makes more localization errors but is far less likely to predict false detections where nothing exists." [source]
Since the release of the first model and possibly due to its success, Redmon and Farhadi (2017) developed an enhanced (stronger and faster) YOLO, but I'll leave that to another day. Until then, if you want to read the full paper, please follow the link below.
As Professor Ng cautioned us, I'll also caution that this is a very technical paper and it can be a challenging read even to the practitioner or the expert in the field. So, take your time...
To stay in touch with me, follow @cristi
Cristi Vlad Self-Experimenter and Author