Paper Reading in 1/1/2016~1/3/2016

Cvpr15:person Count Localization in Videos from Noisy Foreground and detections

Paper the main contribution is to define the person count localization and its perimeter, although the usual result evaluation criteria for the person count problem that was previously mentioned are only the counts given at the end, But in fact, the previous article is not entirely only given to global counts. It is possible to attach more importance to this localization problem and indeed to use this information to solve the problem, but it is necessary to make a distinction, that is, to define the problem first.

This article presents the person count localization problem, which is given in the video sequence, giving the detection and its counts on each frame, as shown in:

After reading it quickly found that, in fact, with the idea of MOT to solve. "Although the article says that this is middle-ground between frame-level person counting and the person detection, but I feel more like this is a rough MOT, at least error driven Graph Revision part of the train of thought is consistent "do not know what kind of effect on the straight running MOT

In fact, at the beginning of the investigation I thought about the method, but I do not know why it is very repulsive tracking by detection method to solve the person count method, may feel that if you do tracking focus on doing tracking it ~ The previous code test, while relentlessly hitting my faith in state of art, is still full of hope.

To get to the next step, paper's approach is to run the person detector and foreground segmentation on the video now, and the two results complement each other. Then using the both the person detector's results and the relatively larger connected components from the foreground segmentation Build the flow graph below, U is Detection,e is the edge of the interconnect, in order to more like a graph, and the same u into two parts, so that the middle also has a side e connection. It satisfies the construction limit of the graph, there will be a traffic x on each edge, inflow equals outflow

Then turn the np-complete problem into a solution ILP problem:

The middle step first omitted, the article several positions have not understood, for example the Scarlet Letter part why request and?

The last is to use the idea of MOT add edge, add node or use tracker to fill up missing detection, and three kinds of operation is to train a random forest classifier, omit ~ So the original flow diagram will change, Repeat these two steps again until the iteration stop condition is met ~

The results of the test because he joined the location information, the traditional method is not applicable, so put forward their own evaluation method, there is nothing to say ~ but at this time back to the previous question, that is, the above-mentioned distinction, if compared with the results of MOT?

Paper Reading in 1/1/2016~1/3/2016