High-precision full-state pose estimation of objects by embedded ultra-compact meta-sensor

# High-precision full-state pose estimation of objects by embedded ultra-compact meta-sensor ## Description We introduce a convolutional neural network for the meta-sensor to intelligently analyzes the shape features of the captured spots, and to efficiently retrieve the embedded positional and orientational information. ## Requirements This work is conducted on a desktop system running Windows 10. The neural network requires Python 3.8.0 or later. All dependencies are listed in the `requirements.txt` file. ```bash# Require python3 and Pytorch>=2.4.1$ pip install -r ./requirements.txt``` ## Quick test Make sure the demo image is placed in the `.\test_images\demo\` folder.Data and model paths can be modified in `testpos_everyin.py`. ```bash# Test CNN$ python testpos_everyin.py``` The results can be found in `.\test_restored\`. ## Train and Evaluate on the demo dataset 1. Train and validate using a subset of the dataset we provide. The dataset directory should follow the basic structure below: ```bashrcDATASET├── dataset| └──train_demo (.jpg)└── groundtruth └──train_demo_gt.csv``` 2. Edit `metavlc.py` to configure settings. ```bashrcfiles_dir = 'dataset/train_demo/'gt=np.genfromtxt('groundtruth/train_demo_gt.csv', delimiter=',', skip_header=0)``` 3. Train the model and evaluate the output results. ```bashpython metavlc.pypython testpos_everyin.py``` ## LicenseThe code is made available for academic research purpose only. Under Attribution-NonCommercial 4.0 International License. ## Citation When using this code, please cite: ```txtHigh-precision full-state pose estimation of objects by embedded ultra-compact meta-sensorMuyang Li, Zhongyi Yuan, Ji ChenSoutheast UniversityPurple Mountain Laboratories```