A TensorFlow implementation of the IMG-Siam tracker

This is a TensorFlow implementation of Initial Matting-Guided Visual Tracking with Siamese Network. The code is improved on the TensorFlow version of SiamFC here.

You can use TensorFlow > 1.0 for tracking though. Note the tracking performance slightly varies in different versions.

# pip install tensorflow    # For CPU
pip install tensorflow-gpu  # For GPU

# Other main modules
pip install sacred, scipy, opencv-python

# Matting needs
pip install scikit-image, scikit-learn, vlfeat-ctypes

# (OPTIONAL) Install nvidia-ml-py for automatically selecting GPU
pip install nvidia-ml-py

git clone https://github.com/lazyfan/IMG-Siam.git

In the initialization phase of the tracker, matting is performed on the initial frame.

You can place the sequence you want to track in the assets, where the sequence video is placed for reference.

(OPTIONAL) There are three matting programs available: sbbm, lbdm, ocsvm, you can modify it in configuration.py

python run_IMGSiam_tracker.py

On the basis of SiamFC, attention module is added to the model, named SiamAtt in paper. The training steps are as follows:

Download and unzip the ImageNet VID 2015 dataset (~86GB) here.

python scripts/preprocess_VID_data.py
# Split train/val dataset and store corresponding image paths
python scripts/build_VID2015_imdb.py

(OPTIONAL) There are two attention modules available: se_block, cbam_block, you can modify it in configuration.py, se_block by default.

python train_SiamAtt.py

Download pretrained models.

python scripts/download_assets.py

Convert pretrained MatConvNet model into TensorFlow format.

# Note we use SiamFC-3s-color-pretrained as one example. You
# Can also use SiamFC-3s-gray-pretrained. 
python convert_pretrained_model.py

Modify trainable variable scope in train_SiamAtt.py and start train.

python train_SiamAtt.py

# Open a new terminal session and cd to IMG-Siam, then
tensorboard --logdir=Logs/track_model_checkpoints/IMGSiam-3s-color

[1] Fully-Convolutional Siamese Networks for Object Tracking

[2] Squeeze-and-Excitation Networks

[1] SiamFC tensorflow implementation

[2] initialisation-problem

[3] CBAM-TensorFlow-Slim