A TensorFlow implementation of the SiamFC tracker
This is a TensorFlow implementation of Fully-Convolutional Siamese Networks for Object Tracking. You can find the original MatConvNet version here. The SiamFC authors have also released a TensorFlow port of the tracking part (using pretrained model only) in here.
This TensorFlow implementation is designed with these goals:
- Self-contained. Data preparation, model training, tracking, visualization and logging all in one.
- Correctness. The tracking performance should be similar to the MatConvNet version.
- Efficiency. The training and inference procedures should be as efficient as possible.
- Modularization. The whole system should be modularized and easy to expand with new ideas.
- Readability. The code should be clear and consistent.
In a computer with a GeForce GTX 1080 GPU, the main results on OTB-100 are in the table below. Note that the performance of SiamFC-3s-color trained from scratch using our implementation is consistently better than SiamFC-3s-gray-scratch. Moreover, We observe that the tracking performance of saved models in different epochs varies considerably, therefore, you may want to evaluate a few more models instead of just picking the model in the last epoch.
| Reported (AUC) | Pretrained (AUC) | Train from scratch (AUC) | Training time | Tracking time | |
|---|---|---|---|---|---|
| SiamFC-3s-color | -- | 0.578 | 0.584 | ~5h | ~120fps |
| SiamFC-3s-gray | 0.582 | 0.587 | 0.573 | ~5h | ~120fps |
Qualitative results:
The main requirements can be installed by:
# (OPTIONAL) 0. It is highly recommended to create a virtualenv or conda environment
# For example,
# conda create -n tensorflow1.4 python=2.7
# source activate tensorflow1.4
# 1. Install TensorFlow 1.4.0
# Version 1.4.0 is required for training since we use tf.data API
# 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
# 2. Install scipy for loading mat files
pip install scipy
# 3. Install sacred for experiments logging
pip install sacred
# 4. Install matplotlib for visualizing tracking results
pip install matplotlib
# 5. Install opencv for preprocessing training examples
pip install opencv-python
# 6. Install pillow for some image-related operations
pip install pillow
# (OPTIONAL) 7. Install nvidia-ml-py for automatically selecting GPU
pip install nvidia-ml-py# 1. Clone this repository to your disk
git clone https://github.com/bilylee/SiamFC-TensorFlow.git
# 2. Change working directory
cd SiamFC-TensorFlow
# 3. Download pretrained models and one test sequence
python scripts/download_assets.py
# 4. 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 experiments/SiamFC-3s-color-pretrained.py
# 5. Run tracking on the test sequence with the converted model
python scripts/run_tracking.py
# 6. Show tracking results
# You can press Enter to toggle between play and pause, and drag the
# scrolling bar in the figure. For more details, see utils/videofig.py
python scripts/show_tracking.py# 1. Download and unzip the ImageNet VID 2015 dataset (~86GB)
# Now, we assume it is unzipped in /path/to/ILSVRC2015
DATASET=/path/to/ILSVRC2015
# 2. Clone this repository to your disk
# (Skip this step if you have already done it in the Tracking section)
git clone https://github.com/bilylee/SiamFC-TensorFlow.git
# 3. Change working directory
cd SiamFC-TensorFlow
# 4. Create a soft link pointing to the ImageNet VID dataset
mkdir -p data
ln -s $DATASET data/ILSVRC2015
# 5. Prepare training data
# If you have followed the data preparation procedures in
# the MatConvNet implementation, simply create a soft link
# pointing to the curated dataset:
# ln -s $CURATED_DATASET data/ILSVRC2015-VID-Curation
# Otherwise, create it from scratch by
python scripts/preprocess_VID_data.py
# 6. Split train/val dataset and store corresponding image paths
python scripts/build_VID2015_imdb.py
# 7. Start training
# You can get quite good results after ~70k iterations.
python experiments/SiamFC-3s-color-scratch.py
# 8. (OPTIONAL) View the training progress in TensorBoard
# Open a new terminal session and cd to SiamFC-TensorFlow, then
tensorboard --logdir=Logs/SiamFC/track_model_checkpoints/SiamFC-3s-color-scratchExample TensorBoard outputs are like:
You can use the run_SiamFC.py in benchmarks directory to integrate with the OTB evaluation toolkit. The OTB toolkit has been ported to python. The original version is here. However, you may want to use my custom version where several bugs are fixed.
Assume that you have followed the steps in Tracking or Training section and now have a pretrained/trained-from-scratch model to evaluate. To integrate with the evaluation toolkit,
# Let's follow this directory structure
# Your-Workspace-Directory
# |- SiamFC-TensorFlow
# |- tracker_benchmark
# |- ...
# 0. Go to your workspace directory
cd /path/to/your/workspace
# 1. Download the OTB toolkit
git clone https://github.com/bilylee/tracker_benchmark.git
# 2. Modify line 22 and 25 in SiamFC-TensorFlow/benchmarks/run_SiamFC.py accordingly.
# In Linux, you can simply run
sed -i "s+/path/to/SiamFC-TensorFlow+`realpath SiamFC-TensorFlow`+g" SiamFC-TensorFlow/benchmarks/run_SiamFC.py
# 3. Copy run_SiamFC.py to the evaluation toolkit
cp SiamFC-TensorFlow/benchmarks/run_SiamFC.py tracker_benchmark/scripts/bscripts
# 4. Add the tracker to the evaluation toolkit list
echo "\nfrom run_SiamFC import *" >> tracker_benchmark/scripts/bscripts/__init__.py
# 5. Create tracker directory in the evaluation toolkit
mkdir tracker_benchmark/trackers/SiamFC
# 6. Start evaluation (it will take some time to download test sequences).
echo "tb100" | python tracker_benchmark/run_trackers.py -t SiamFC -s tb100 -e OPE
# 7. Get the AUC score
sed -i "s+tb50+tb100+g" tracker_benchmark/draw_graph.py
python tracker_benchmark/draw_graph.pySiamFC-TensorFlow is released under the MIT License (refer to the LICENSE file for details).