search

amoudgl / mosse-tracker

Implementation of MOSSE tracker in MATLAB: Visual Object Tracking using Adaptive Correlation Filters (CVPR 2010)
MIT License
82 stars 25 forks source link

train N template for each frame, and del rotate #3

Closed happy-v587 closed 5 years ago

happy-v587 commented 5 years ago

通过阅读作者的论文，发现更新模板是按照如下策略进行： _20181109203612 H本身定义如下： _20181109203650 我认为作者的本意是：每一帧都进行N次采样和训练。但由于作者把N个采样的下标也用的i，使得下标i和第i帧重复使用了

amoudgl commented 5 years ago

@HappyUncle As far as I know, MOSSE tracker does not do random affine transformation for each frame as reflected in your pull request. Authors mention this fact in section 3.4 of the paper:

The training set is constructed using random affine transformations to generate eight small perturbations (fi) of the tracking window in the initial frame. Training outputs (gi) are also generated with their peaks corresponding to the target center. During tracking, a target can often change appearance by changing its rotation, scale, pose, by moving through different lighting conditions, or even by undergoing nonrigid deformation. Therefore, filters need to quickly adapt in order to follow objects. Running average is used for this purpose.

You can also verify this from OpenCV implementation of MOSSE.

happy-v587 commented 5 years ago

okay I think it would be better to randomly extract some samples for each frame.

amoudgl commented 5 years ago

Yes, incremental improvements are definitely possible but in this repository, I want to keep the code which follows the paper closely.

Anyway, thanks for your comments! I'll go ahead and close this for now :)