In previous post about generic object recognition, identity report is briefly mentioned. But as current generic object recognition doesn’t focus on this problem, no further explanation about it.
This post will talk about currently most popular identity system, face recognition.
This series posts briefly explained object recognition and some popular deep learning based models.
Continue reading Object Recognition (8) â Summary
Though YOLO improves the detection speed drastically, it sacrifices accuracy. The accuracy is lower than Faster R-CNN.
YOLO is faster because it doesnât use proposal. To keep the speed and improve accuracy, a new method needed to compensate the lacking of proposal. And the method should be light weight.
In previous posts, we talked about the R-CNN series models for object recognition. We see how it evolves from multi-stage architecture to single stage single network. Now we know the object detection model can be simplified to just a single network. Can we make things more simpler by dropping the RPN and just use a deep CNN network? YOLO answer this question with yes.
Mask R-CNN is the last model of R-CNN series.
As the introduction, object location can be reported roughly as bounding box, or finely as pixels. After optimizing the R-CNN to good accuracy of both category and location detection, now the next step is to try report pixels for object location instead of bounding box.
Continue reading Object Recognition (5) â Mask R-CNN
Faster R-CNN is the fastest model of the R-CNN series.As the name suggested, itâs faster than the Fast R-CNN.
Faster R-CNN solved the major remain problem of Fast R-CNN. The Selective Search is replaced with Region Proposal Network(RPN) and merged into the main network. Now the detection only require a single pass through the network.
Continue reading Object Recognition (4) â Faster R-CNN
Last post give a brief introduction to R-CNN, also mentioned some problems in it. This post will discuss the successor, Fast R-CNN.
Continue reading Object Recognition (3) â Fast R-CNN
R-CNN is the first of a series important models archive impressing results. Following R-CNN, fast R-CNN, faster R-CNN improve on both accuracy and efficiency. And Mask R-CNN even improve the model from using bounding box for location to using mask for location.
Object recognition is one of the hottest areas of Artificial Intelligence.
The ultimate target is the visual ability like human being. That is recognizing the scene, every object and background. So for object recognition, computer should report every objectâs location and category. If the any object is the known specific one, identity should also be reported.
Continue reading Object Recognition (1) â Introduction
Recently some new security problems have been found with modern computer processors with speculative execution. After reading blog posts from google, some papers and mailing list, I got a rough understanding of these problems and the possible mitigations. (936) 463-5740