rockkingjy / eco Goto Github PK

This package requires matconvnet [1], if you want to use deep CNN features, and PDollar Toolbox [2], if you want to use HOG features.

Download matconvnet ZIP file from https://github.com/vlfeat/matconvnet and unpack it in the external_libs/matconvnet/ folder of the repository.

Download PDollar Toolbox ZIP file from https://github.com/pdollar/toolbox and unpack it in the external_libs/pdollar_toolbox/ folder of the repository.

install
demo_ECO

The files in the runfiles/ directory are uset to set parameters and run the tracker. You can create your own runfiles by copying an existing one and then play around with different parameters and feature combinations.

These runfiles are included:

• OTB_DEEP_settings.m - Settings in the paper used for the deep feature version of ECO on the OTB, UAV123 and TempleColor datasets.

• OTB_HC_settings.m - Settings in the paper used for the hand-crafted feature version of ECO (ECO-HC) on the OTB, UAV123 and TempleColor datasets.

• VOT2016_DEEP_settings.m - Settings in the paper used for the deep feature version of ECO on the VOT2016 dataset.

• VOT2016_HC_settings.m - Settings in the paper used for the hand-crafted feature version of ECO (ECO-HC) on the VOT2016 dataset.

• testing_ECO.m - Demo file which contains the same settings as OTB_DEEP_settings.m by default.

• testing_ECO_gpu.m - Same as testing_ECO, but uses GPU.

• testing_ECO_HC.m - Demo file which contains the same settings as OTB_HC_settings.m by default.

• SRDCF_settings.m - Settings that roughly correspond to our previous SRDCF tracker [4] from 2015. The main difference is the optimization method.

Tracking performance on the OTB-2015 dataset is shown bellow for different settings. For comparison, results of our previous trackers C-COT [3], SRDCF [4], DeepSRDCF [5] and DSST [6] are included.

This package includes a quite general framework for feature extraction. You can easily incorporate your own features in the same manner by adding a corresponding "get_featureX.m" function.

Currently, four types of features are included:

1. Deep CNN features. It uses matconvnet [1], which is included as a git submodule in external_libs/matconvnet/. The imagenet-vgg-m-2048 network available at http://www.vlfeat.org/matconvnet/pretrained/ was used in the ECCV paper. You can use other networks, by placing them in the feature_extraction/networks/ folder.

2. HOG features. It uses the PDollar Toolbox [2], which is included as a git submodule in external_libs/pdollar_toolbox/.

3. Lookup table features. These are implemented as a lookup table that directly maps an RGB or grayscale value to a feature vector. Currently, Color Names [8] and Intensity Channels [9] are included.

4. Colorspace features. Currently grayscale and RGB are implemented.

The tracker supports almost any combination of features. Currently, the only limitation is that you can only use deep features from a single network (but you can use several different layers from the same network).

Each feature has its own parameter settings. You can set the cell size for each non-CNN feature independently. ECO does not assume the same cell size for all feature channels. For the CNN features, you can control the cell size by setting an additional down-sampling factor for each layer.

See the runfile testing.m for examples of how to integrate different features. You can uncomment several features at once in the params.t_features cell array.

This reposetery also includes an implementation of the optimized scale filter, which was originally proposed in the fDSST [7]. It is here used in the ECO-HC version of the tracker for speeding-up the scale estimation.

GPU support is activated by setting the parameter "params.use_gpu = true" in the runfile. This requires MatConvNet to be compiled with GPU support. If the install script fails, please visit http://www.vlfeat.org/matconvnet/install/ for instructions.

It should be easy to integrate the tracker into the Online Tracking Benchmark [10]. The runfiles supports the OTB interface, so you just have to copy and rename the runfile you want to use and then add the necessary paths (see setup_paths.m).

To integrate the tracker into the Visual Object Tracking (VOT) challenge toolkit [11], check the VOT_integration folder. Copy the configuration file to your VOT workspace and set the path to the ECO reposetory inside it. The tracker now supports the trax protocol, which is necessary for VOT2017 version of the toolkit.

All raw result files used in our CVPR 2017 paper can be found at the project webpage: http://www.cvl.isy.liu.se/research/objrec/visualtracking/ecotrack/index.html

Tracking performance may vary slightly on different machines and whether GPU support is activated. This is due to small numerical effects which can accumulate over time (all trackers are affected by this). Generally, this only affects the performance marginally. More significant changes are rare, but can occur in some videos.

[1] Webpage: http://www.vlfeat.org/matconvnet/
GitHub: https://github.com/vlfeat/matconvnet

[2] Piotr Dollár.
"Piotr’s Image and Video Matlab Toolbox (PMT)."
Webpage: https://pdollar.github.io/toolbox/
GitHub: https://github.com/pdollar/toolbox

[3] Martin Danelljan, Andreas Robinson, Fahad Khan, Michael Felsberg.
Beyond Correlation Filters: Learning Continuous Convolution Operators for Visual Tracking.
In Proceedings of the European Conference on Computer Vision (ECCV), 2016.
http://www.cvl.isy.liu.se/research/objrec/visualtracking/conttrack/index.html

[4] Martin Danelljan, Gustav Häger, Fahad Khan, Michael Felsberg.
Learning Spatially Regularized Correlation Filters for Visual Tracking.
In Proceedings of the International Conference in Computer Vision (ICCV), 2015.
http://www.cvl.isy.liu.se/research/objrec/visualtracking/regvistrack/index.html

[5] Martin Danelljan, Gustav Häger, Fahad Khan, Michael Felsberg.
Convolutional Features for Correlation Filter Based Visual Tracking.
ICCV workshop on the Visual Object Tracking (VOT) Challenge, 2015.
http://www.cvl.isy.liu.se/research/objrec/visualtracking/regvistrack/index.html

[6] Martin Danelljan, Gustav Häger, Fahad Khan and Michael Felsberg.
Accurate Scale Estimation for Robust Visual Tracking.
In Proceedings of the British Machine Vision Conference (BMVC), 2014.
http://www.cvl.isy.liu.se/research/objrec/visualtracking/scalvistrack/index.html

[7] Martin Danelljan, Gustav Häger, Fahad Khan, Michael Felsberg.
Discriminative Scale Space Tracking.
Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2017.
http://www.cvl.isy.liu.se/research/objrec/visualtracking/scalvistrack/index.html

[8] J. van de Weijer, C. Schmid, J. J. Verbeek, and D. Larlus.
Learning color names for real-world applications.
TIP, 18(7):1512–1524, 2009.

[9] M. Felsberg.
Enhanced distribution field tracking using channel representations.
In ICCV Workshop, 2013.

[10] Y. Wu, J. Lim, and M.-H. Yang.
Online object tracking: A benchmark.
In CVPR, 2013.
https://sites.google.com/site/trackerbenchmark/benchmarks/v10

[11] http://votchallenge.net/