Homography Decomposition Networks for Planar Object Tracking

This project is the mindspore version of HDN(Homography Decomposition Networks for Planar Object Tracking) , this paper was accepted by AAAI 2022.

Project Page | Paper| Pytorch Version Code

MindSpore

MindSpore is a deep learning framework in all scenarios, aiming to achieve easy development, efficient execution, and all-scenario coverage. Please check the official homepage.

Installation

To strat, install MindSpore. Please find python dependencies and installation instructions in INSTALL.md.

The code is tested on an Ubuntu 18.04 system with Nvidia GPU RTX 3090Ti.

Requirments

• Conda with Python 3.7
• Nvidia GPU
• MindSpore >= 1.8.0
• pyyaml
• yacs
• tqdm
• matplotlib
• OpenCV
• ....

Quick Start

Add HDN_mindspore to your PYTHONPATH

export PYTHONNPATH=/path_to_HDN_mindspore:/path_to_HDN_mindspore/homo_estimator/Deep_homography/Oneline_DLTv2:$PYTHONPATH

Download models

In the pretrained_models, download the pretrained weights.Baidu Netdisk key: f9K7

Config and Datasets

For the global parameters and datasets, please refer to the original project Readme.

Test

cd experiments/tracker_homo_config/
python ../../tools/test.py --snapshot ../../hdn.ckpt --config proj_e2e_GOT_unconstrained_v2.yaml --dataset POT210 --video --vis

To test multiple datasets it is recommended to use muti_test:

python ../../tools/muti_test.py

The test accuracy is basically up to the standard in POT201. (you need to run multiple times):

TODO List