Object tracking papers (1): SINT--Siamese Instance Search for Tracking

Time: January 24th, 2018, the ninteenth week

Weekly Summary: During this week, I sequentially read a paper about employing instance search in VOT.

Paper: Siamese Instance Search for tracking/ Author: Ran Tao, Efstratios Gavves, Arnold W.M. Smeulders/ Publication information:CVPR 2016



Outline:
We call this work SINT for abbreviation. It is very different from the paper I read before. It utilizes the Siamese Network to train the matching function between the groundtruth in one frame and the candidates in other frames. In this case the network gets
a matching function to adapt all kinds of distortions. It didn't learn the features of the target,but only the variation of targets. And it contrasts the AlexNet and the VGGNet. They drop the max pooling layer in AlexNet, and keep the early two layers in VGGNet,
because the max pooling layer is not very helpful to localize precisely. And it adds overlapping region layers to  decrease the burden of learning, using one image and its relevant regions as input. And it normalizes each layer activations and the input of
the loss function to let the value in a certain sphere.  Certainly the training data should be totally different from the test videos, and the target should not be the same. During tracking process, there is no updating model or learning online. The algorithm
always finds the best matching in each frame to the target in the first frame. The experiments of this work is very sufficient. First they compare the algorithms with different network and different training methods. And then it is compared to other state-of-the-art
algorithms.



Advantages: They deploy the way in instance search into object tracking. And it does not force the trackers to deal with specific distortions. And it avoids overfitting in tracking, somehow.

Disadvantages: They just test the methods in OTB50. It's not very sufficient to verify the accuracy and robustness of the SINT tracker. The training data is not very large to distinguish it from the other trackers. And the inner comparison
is not enough to prove the drawbacks of maxpooling and the benefits of VGGNet.