[target tracking] |dimp: learning discriminative model prediction for tracking

The paper ：DIMP: Learning Discriminative Model Prediction for Tracking
Code ：https://github.com/visionml/pytracking

Send the target area into the feature extractor to get a better feature representation . Siamese tracking The algorithm only uses the target area , Features to be extracted in the target area , Directly perform relevant operations in the features of the test frame image , This method is simple , Fast , The effect is good in simple scenes , But the robustness is poor , Unable to deal with A complex scenario （ Similar object problem ）. There is no online classifier

To solve the problem siamese The network is not robust enough , Propose a new framework ,

Use multi frame images for training , Use random samples in the video sequence for training （ Extract three frames from a certain frame and the front as training set, Three frames are extracted from the back of this frame as a test set）