A Python+Tensorflow implementation of siamese-fc
This is the Python+Tensorflow implementation of Fully-Convolutional Siamese Networks for Object Tracking, including both training and tracking.
The original Matlab version can be found at siamese-fc.
The tracker borrows a lot of code from py-siamese_fc, many thanks.
ref: Fully-Convolutional Siamese Networks for Object Tracking
Prerequisites:
python: 3.4; tensorflow: 1.0.1
-
[ Tracking only ] For tracking, pretrained networks as ckpt file can be plugged into the
tracker.py
directly.- Clone the repository.
- A pretrained networks with Imagenet VID dataset can be downloaded from baidu pan, unzip the file to
./models/
. - Execute
tracker.py
, the video sequence (./demo-sequences/vot15_bag
) is processed as an example.
-
[ Training ] To train a model, following steps can be considered.
- Clone the reposistory.
- Follow the instructions from original version to generate the curated dataset for training.
- Open the created
imageStats.mat
with MATLAB, upzipx.mat
andz.mat
in the path./ILSVRC15-curation/
, runcuration.py
to getimdb.pkl
andimageStats.pkl
as python version ofimdb_video.mat
andimageStats.mat
. - Execute
train.py
to train your own model. Tensorboard can also be used during this phase to monitor variables in the network. - Parameters of networks are saved as ckpt files in
./ckpt
for tracking.