https://github.com/zj5559/DATr

PyTorch implementation of "Leveraging the Power of Data Augmentation for Transformer-based Tracking" (WACV2024)
https://github.com/zj5559/DATr

Last synced: 6 days ago
JSON representation

PyTorch implementation of "Leveraging the Power of Data Augmentation for Transformer-based Tracking" (WACV2024)

• Host: GitHub
• URL: https://github.com/zj5559/DATr
• Owner: zj5559
• License: mit
• Created: 2023-09-15T07:27:39.000Z (about 1 year ago)
• Default Branch: master
• Last Pushed: 2023-11-14T02:04:43.000Z (12 months ago)
• Last Synced: 2024-06-17T00:27:23.824Z (5 months ago)
• Language: Python
• Homepage:
• Size: 1.45 MB
• Stars: 9
• Watchers: 2
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• Awesome-Visual-Object-Tracking - [code

README

        
# DATr 

PyTorch implementation of "Leveraging the Power of Data Augmentation for Transformer-based Tracking" (WACV2024).

Please find the paper [here](https://arxiv.org/pdf/2309.08264.pdf).

## Introduction

In this paper, we perform systematic experiments to explore the impact of General Data Augmentations (GDA) on transformer trackers, including the pure transformer tracker and the hybrid CNN-Transformer tracker. Results below show GDAs have limited effects on SOTA trackers.

![DATR figure](experiments.png)

Then, We propose two Data Augmentation methods based on challenges faced by Transformer-based trackers, DATr for short. They improve trackers from perspectives of adaptability to different scales, flexibility to boundary targets, and robustness to interference, respectively.

![DATR figure](framework.png)

Extensive experiments on different baseline trackers and benchmarks demonstrate the effectiveness and generalization of our DATr, especially for sequences with challenges and unseen classes.

![DATR figure](results.png)

## Installation

The environment installation and training configurations (like project path, pretrained models) are similar to the baseline trackers, e.g., OSTrack, please refer to [OSTrack](https://github.com/botaoye/OSTrack). 

## Training and Testing

Please see eval.sh to find the commands for training and testing.

## Models and Results

Models and results can be found [here](https://drive.google.com/drive/folders/19-jBvfFVZxPcvZmy6ZXwtyW6NCnQ6bjY?usp=share_link).

## Acknowledgments

Our work is mainly implemented on three different Transformer trackers, i.e., [OSTrack](https://github.com/botaoye/OSTrack), [MixFormer](https://github.com/MCG-NJU/MixFormer), and [STARK](https://github.com/MasterBin-IIAU/Stark-1). Thanks for these concise and effective SOT frameworks.