SiamRPN - PyTorch

A clean PyTorch implementation of SiamRPN tracker described in paper High Performance Visual Tracking with Siamese Region Proposal Network. The code is evaluated on 7 tracking datasets (OTB (2013/2015), VOT (2018), DTB70, TColor128, NfS and UAV123), using the GOT-10k toolkit.

Performance

GOT-10k

Dataset	AO	SR0.50	SR0.75
GOT-10k	0.462	0.556	0.218

The scores surpass the highest performance on GOT-10k leaderboard (AO 0.374, SR_0.50 0.404) by a large margin.

However, since SiamRPN is trained on 4 extra datasets (ILSVRC-VID, YouTube-BB, ImageNet Detection and COCO) and it does not follow the one-shot principle (zero-overlap between training and test object classes) of GOT-10k, the comparison may not be fair.

OTB / UAV123 / DTB70 / TColor128 / NfS

Dataset	Success Score	Precision Score
OTB2013	0.641	0.855
OTB2015	0.629	0.837
UAV123	0.599	0.770
UAV20L	0.531	0.656
DTB70	0.548	0.756
TColor128	0.533	0.736
NfS (30 fps)	0.453	0.529
NfS (240 fps)	0.589	0.706

VOT2018

Dataset	Accuracy	Robustness (unnormalized)
VOT2018	0.576	27.00

Dependencies

Install PyTorch, opencv-python and GOT-10k toolkit:

pip install torch
pip install opencv-python
pip install --upgrade git+https://github.com/got-10k/toolkit.git@master

GOT-10k toolkit is a visual tracking toolkit that implements evaluation metrics and tracking pipelines for 7 main datasets (GOT-10k, OTB, VOT, UAV123, NfS, etc.).

Running the tracker

In the root directory of siamrpn-pytorch:

1. Download pretrained model.pth from Baidu Yun or Google Drive, and put the file under pretrained/siamrpn.

2. Create a symbolic link data to your datasets folder (e.g., data/OTB, data/UAV123, data/GOT-10k).

3. Run:

python run_tracking.py

By default, the tracking experiments will be executed and evaluated over all 7 datasets. Comment lines in run_tracker.py as you wish if you need to skip some experiments.