Towards Accurate Readings of Water Meters by Eliminating Transition Error: New Dataset and Effective Solution

This is the official implementation of our paper Towards Accurate Readings of Water Meters by Eliminating Transition Error: New Dataset and Effective Solution.

News

🔥 A comprehensive list of Awesome Image-based Meter Recognition Reading methods and datasets is available.

Methodology

WMeter5K dataset

WMeter5K totally contains 5,000 water meter images, which are all captured by the add-on camera attached to real-world water meters. Fine-grained annotations are provided for each image, including bounding box and readings for each digital wheel and pointer, and the overall readings of the entire meter.

1. Down load WMeter5K from here and put it to ./dataset/WMeter5K
2. Use ./dataset/visualize.py for visualization.

Inference

1. Put the model weights im.pkl and tr.pkl to ./checkpoint/
2. Prepare the water meter image and the corresponding detection annotations. Use ./demo/cropping.py to obtain cropped pointer images.
3. Run the following script for inference

python val_and_test.py --mode test --img_folder ./demo/pointers/

Evaluation

1. Down load WMeter5K from here and put it to ./dataset/WMeter5K
2. Run the following script for evaluation on WMeter5K's test set

python val_and_test.py --mode val

Training

1. Down load WMeter5K from here and put it to ./dataset/WMeter5K
2. Prepare the LMDB version for WMeter5K by runing

python createlmdb.py

3. Put the model weights im.pkl to ./checkpoint/
4. Run the following script for training

python train.py

Citation:

@inproceedings{zhangwater2024, 
Author = {Jiaxin Zhang, Daizhen Jia, Chongyu Liu, Dezhi Peng, Bangdong Chen, Xue Gao and Lianwen Jin}, 
Booktitle = {IEEE Transactions on Instrument and Measurement}, 
Title = {Towards Accurate Readings of Water Meters by Eliminating Transition Error: New Dataset and Effective Solution, 
Year = {2024}}   

⭐ Star Rising