How hard can it be? Estimating the difficulty of visual search in an image and PPT

How hard can it be? Estimating the difficulty of visual search in an image published in 2016 CVPR, this article is recommended by the teacher, it should be explained that its model is relatively simple, and its real idea is relatively new.

The main work of the author is to propose a method to measure the difficulty of searching the image in the visual search. We know that people's vision in the search, the first to look at the search image, to obtain the image of the important subject of the characteristics, and then search for similar images, that is, the first image abstraction, then search. In the current study, most people are studying how to improve the accuracy of search, but the author is thinking, what kind of images are easy to retrieve, what is difficult to retrieve. A metaphor, there is a rope, some people say 1 meters, some people say 2 meters, in the end is how much, we have to take a ruler to measure, the method proposed by the author is equivalent to "ruler", to solve a measurement problem, similar to the Qin Shihuang unified weights and measures, we will take this to measure, so this is a pioneering work.

as a pioneering work, the data generally need to get their own, so the first part of the author also spent a lot of energy to mark the data, the use of the response time as a marker, then the author proves that the search difficulty of the image is proportional to the response time of the person, that is, the time tag is reasonable And then the author analyzes the seven factors that affect the difficulty of image search from the perspective of human cognition, and analyzes the degree of influence respectively; then the author presents his own model. CNN + V-SVR and image difficulty score calculation method: The geometric median of all marker time of the picture (square root of time multiplication) as difficulty score ; Compared with baselines, the result is better; Finally, the author proposes that the model can also play a role in the field of weak supervision target localization and semi-supervised target classification, and gives the experimental proof. For more information, please see another blog:http://blog.csdn.net/helloeveryon/article/details/51893251

The model is simple, but the idea is worth learning. because the blog can not upload ppt, so I directly map. Students who need PPT can leave a message.

How hard can it be? Estimating the difficulty of visual search in an image and PPT