Awesome

SiamFC-PyTorch

This is the PyTorch (0.40) implementation of SiamFC tracker [1], which was originally <a href="https://github.com/bertinetto/siamese-fc">implemented</a> in Matlab using MatConvNet [2]. In our implementation, we obtain better perforamnce than the original one.

Goal

• A more compact implementation of SiamFC [1]
• Reproduce the results of SiamFC [1], including data generation, training and tracking

Requirements

• Python 2.7 (I use Anaconda 2.* here. If you use Python3, you may get the very different results!)
• Python-opencv
• PyTorch 0.40
• other common packages such as numpy, etc

Data curation

• Download <a href="http://bvisionweb1.cs.unc.edu/ilsvrc2015/ILSVRC2015_VID.tar.gz">ILSVRC15</a>, and unzip it (let's assume that $ILSVRC2015_Root is the path to your ILSVRC2015)

• Move $ILSVRC2015_Root/Data/VID/val into $ILSVRC2015_Root/Data/VID/train/, so we have five sub-folders in $ILSVRC2015_Root/Data/VID/train/

• It is a good idea to change the names of five sub-folders in $ILSVRC2015_Root/Data/VID/train/ to a, b, c, d, and e

• Move $ILSVRC2015_Root/Annotations/VID/val into $ILSVRC2015_Root/Annotations/VID/train/, so we have five sub-folders in $ILSVRC2015_Root/Annotations/VID/train/

• Change the names of five sub-folders in $ILSVRC2015_Root/Annotations/VID/train/ to a, b, c, d and e, respectively

• Generate image crops

  • cd $SiamFC-PyTorch/ILSVRC15-curation/ (Assume you've downloaded the rep and its path is $SiamFC-PyTorch)
  • change vid_curated_path in gen_image_crops_VID.py to save your crops
  • run $python gen_image_crops_VID.py (I run it in PyCharm), then you can check the cropped images in your saving path (i.e., vid_curated_path)

• Generate imdb for training and validation

  • cd $SiamFC-PyTorch/ILSVRC15-curation/
  • change vid_root_path and vid_curated_path to your custom path in gen_imdb_VID.py
  • run $python gen_imdb_VID.py, then you will get two json files imdb_video_train.json (~ 430MB) and imdb_video_val.json (~ 28MB) in current folder, which are used for training and validation

Train

• cd $SiamFC-PyTorch/Train/
• Change data_dir, train_imdb and val_imdb to your custom <b>cropping path</b>, training and validation json files
• run $python run_Train_SiamFC.py
• <b>some notes in training</b>
  • the parameters for training are in Config.py
  • by default, I use GPU in training, and you can check the details in the function train(data_dir, train_imdb, val_imdb, model_save_path="./model/", use_gpu=True)
  • by default, the trained models will be saved to $SiamFC-PyTorch/Train/model/
  • each epoch (50 in total) may take 7-8 minuts (Nvidia 1080 GPU), and you can use parallelling utilities in PyTorch for speeding up
  • I tried to use fixed random seeds to get the same results, but it doesn't work ):, so results for each training may be slightly different (still better than the original)
  • <b>only</b> color images are used for training, and better performance is expected if using color+gray as in original paper

Test (Tracking)

• cd $SiamFC-PyTorch/Tracking/
• <b>Firstly</b>, you should take a look at Config.py, which contains all parameters for tracking
• Change self.net_base_path to the path saving your trained models
• Change self.seq_base_path to the path storing your test sequences (OTB format, otherwise you need to revise the function load_sequence() in Tracking_Utils.py
• Change self.net to indicate whcih model you want for evaluation (by default, use the last one), and I've uploaded a trained model SiamFC_50_model.pth in this rep (located in $SiamFC-PyTorch/Train/model/)
• Change other parameters as your willing :)
• Now, let's run $python run_Train_SiamFC.py
• <b>some notes in tracking</b>
  • two evaluation types are provided: single video demo and evaluation on the whole (OTB-100) benchmark
  • you can also change whihc net for evaluation in run_Train_SiamFC.py

Results

I tested the trained model on OTB-100 using a Nvidia 1080 GPU. The results and comparisons to the original implementation are shown in the below image. The running speed of our implementation is <b>82 fps</b>. Note that, both models are trained from stratch.

References

[1] L. Bertinetto, J. Valmadre, J. F. Henriques, A. Vedaldi, and P. H. Torr. Fully-convolutional siamese networks for object tracking. In ECCV Workshop, 2016.

[2] A. Vedaldi and K. Lenc. Matconvnet – convolutional neural networks for matlab. In ACM MM, 2015.

Contact

Any question are welcomed to hengfan@temple.edu.