November 4, 2019

414 words 2 mins read



SiamRPN, SiamRPN++, unofficial implementation of “SiamRPN++” (CVPR2019), multi-GPUs, LMDB.

repo name PengBoXiangShang/SiamRPN_plus_plus_PyTorch
repo link
language Python
size (curr.) 766 kB
stars (curr.) 374
created 2019-04-12


This is an unofficial PyTorch implementation of SiamRPN++ (CVPR2019), implemented by Peng Xu and Jin Feng. Our training can be conducted on multi-GPUs, and use LMDB data format to speed up the data loading.

This project is designed with these goals:

  • Training on ILSVRC2015_VID dataset.
  • Training on GOT-10k dataset.
  • Training on YouTube-BoundingBoxes dataset.
  • Evaluate the performance on tracking benchmarks.

Details of SiamRPN++ Network

As stated in the original paper, SiamRPN++ network has three parts, including Backbone Networks, SiamRPN Blocks, and Weighted Fusion Layers.

1. Backbone Network (modified ResNet-50)

As stated in the original paper, SiamRPN++ uses ResNet-50 as backbone by modifying the strides and adding dilated convolutions for conv4 and conv5 blocks. Here, we present the detailed comparison between original ResNet-50 and SiamRPN++ ResNet-50 backbone in following table.

2. SiamRPN Block

Based on our understanding to the original paper, we plot a architecture illustration to describe the Siamese RPN block as shown in following.

We also present the detailed configurations of each layer of RPN block in following table. Please see more details in ./network/

component configuration
adj_1 / adj_2 / adj_3 / adj_4 conv2d(256, 256, ksize=3, pad=1, stride=1), BN2d(256)
fusion_module_1 / fusion_module_2 conv2d(256, 256, ksize=1, pad=0, stride=1), BN2d(256), ReLU
box head conv2d(256, 4*5, ksize=1, pad=0, stride=1)
cls head conv2d(256, 2*5, ksize=1, pad=0, stride=1)

3. Weighted Fusion Layer

We implemente the weighted fusion layer via group convolution operations. Please see details in ./network/


Ubuntu 14.04

Python 2.7

PyTorch 0.4.0

Other main requirements can be installed by:

# 1. Install cv2 package.
conda install opencv

# 2. Install LMDB package.
conda install lmdb

# 3. Install fire package.
pip install fire -c conda-forge

Training Instructions

# 1. Clone this repository to your disk.
git clone

# 2. Change working directory.
cd SiamRPN_plus_plus_PyTorch

# 3. Download training data. In this project, we provide the downloading and preprocessing scripts for ILSVRC2015_VID dataset. Please download ILSVRC2015_VID dataset (86GB). The cripts for other tracking datasets are coming soon.
cd data
wget -c
tar -xvf ILSVRC2015_VID.tar.gz
rm ILSVRC2015_VID.tar.gz
cd ..

# 4. Preprocess data.
chmod u+x ./preprocessing/

# 5. Pack the preprocessed data into LMDB format to accelerate data loading.
chmod u+x ./preprocessing/

# 6. Start the training.
chmod u+x ./


Many thanks to Sisi who helps us to download the huge ILSVRC2015_VID dataset.

comments powered by Disqus