The official implementation of paper Siamese Transformer Pyramid Networks for Real-Time UAV Tracking, accepted by WACV22

Last update: Jan 08, 2023

Related tags

Deep Learning SiamTPNTracker

Overview

SiamTPN

Introduction

This is the official implementation of the SiamTPN (WACV2022). The tracker intergrates pyramid feature network and transformer into Siamese network, achieving state-of-the-art performance (better than DiMP) while runing 30 FPS on a single CPU. The tracker optimized with ONXX and openvino could run at 45 FPS on cpu end, leading promising performance when deploying on drones for tracking.

[Paper] [Raw Results] [Drone Tracking Videos] [Models]

Training

prepare data

change the path in lib/train/admin/local.py to your data location

# Distributed training withh 4 nodes 
python -m torch.distributed.launch --nproc_per_node 4 tools/run_training.py --config shufflenet_l345_192

# single gpu training for test purpose
python tools/run_training.py --config shufflenet_l345_192

Test and evaluate SiamTPN

prepare data

change the path in lib/test/evaluation/local.py to your data location

running on cpu

# Download the pretrain model and put it under ./results/checkpoints/train/SiamTPN/ folder

python tools/test.py siamtpn shufflenet_l345_192 --dataset_name got10k_val --debug 1 --cpu 1 --epoch 100 --sequence GOT-10k_Val_000001

running on cpu with onnx optimized

The debug mode will show tracking results, more details refer to tools/test.py

Currently, onnx only support cpu version

First, you need to install onxx and onxxruningtime:

pip install onxx
# for onxx runining time, download the openvino version from release [page](https://github.com/intel/onnxruntime/releases/tag/v3.1) and install with
pip install onnxruntime_openvino-1.9.0-cp37-cp37m-linux_x86_64.whl

# please refer the [page](https://github.com/intel/onnxruntime/releases/tag/v3.1) for openvino installation details.

# Download the converted onnx model and put it under ./results/onnx/ folder
# or conver your own model with 
python tools/onnx_search.py
python tools/onnx_template.py

python tools/test.py siamtpn_onnx shufflenet_l345_192 --dataset_name got10k_val --debug 1 --cpu 1 --epoch 100 --sequence GOT-10k_Val_000001

Citation

Acknowledge

Our code is implemented based on the following libraries:

The official implementation of paper Siamese Transformer Pyramid Networks for Real-Time UAV Tracking, accepted by WACV22

Related tags

Overview

SiamTPN

Introduction

Training

prepare data

Test and evaluate SiamTPN

prepare data

running on cpu

running on cpu with onnx optimized

Citation

Acknowledge

Owner

Robotics and Intelligent Systems Control @ NYUAD

Multitask Learning Strengthens Adversarial Robustness

QuadTree Attention for Vision Transformers (ICLR2022)

PenguinSpeciesPredictionML - Basic model to predict Penguin species based on beak size and sex.

True per-item rarity for Loot

Simple codebase for flexible neural net training

Automatic deep learning for image classification.

A python module for configuration of block devices

MXNet implementation for: Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution

StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks

Implementation of the paper "Language-agnostic representation learning of source code from structure and context".

Mesh Graphormer is a new transformer-based method for human pose and mesh reconsruction from an input image

Towards Boosting the Accuracy of Non-Latin Scene Text Recognition

A Transformer-Based Feature Segmentation and Region Alignment Method For UAV-View Geo-Localization

Official implementation of the Neurips 2021 paper Searching Parameterized AP Loss for Object Detection.

A PyTorch implementation of "Multi-Scale Contrastive Siamese Networks for Self-Supervised Graph Representation Learning", IJCAI-21

Tools for investing in Python

Official repository of "DeepMIH: Deep Invertible Network for Multiple Image Hiding", TPAMI 2022.

Definition of a business problem according to Wilson Lower Bound Score and Time Based Average Rating

Pytorch code for "Text-Independent Speaker Verification Using 3D Convolutional Neural Networks".

Official pytorch implementation of paper "Inception Convolution with Efficient Dilation Search" (CVPR 2021 Oral).