RebarCounting

A computer vision deep learning model to automatically count the number of rebars, based on https://arxiv.org/abs/1807.09856

Installation

First clone the repository. Run

git clone https://github.com/andohuman/RebarCounting.git

Since the repository also contains some large files (weights) you might additionally want to run

git lfs install
git lfs pull

This will download the weights (for predition)

Then, if you have pip installed, run

pip install -r requirements.txt

This will install all the requirements for the code.

then run

python3 train.py -r __misc/ -l 50 -ne 1

to see if the training script is working right. This should generate a bad prediction in _visualizing_dots/1/

This code requires cuda 10.0 and corresponding cudnn to be installed on the system. Code was ran and tested on python 3.6.8, torch 1.0.1 and torchvision 0.2.2

Usage

For training

Run

python3 train.py -h
usage: train.py [-h] [-r ROOT] [-e EPOCH] [-l LAYERS] [-ne NUM_EPOCH]
                [-se SAVE_EVERY] [-v VISDOM]

Train the counting model

optional arguments:
  -h, --help            show this help message and exit
  -r ROOT, --root ROOT  Root directory housing the data
  -e EPOCH, --epoch EPOCH
                        Epoch to start from (make sure epoch.pth exists in
                        model_data/model and model_data/opt)
  -l LAYERS, --layers LAYERS
                        resnet version to use
  -ne NUM_EPOCH, --num_epoch NUM_EPOCH
                        number of epochs to train the model
  -se SAVE_EVERY, --save_every SAVE_EVERY
                        save every - iterations
  -v VISDOM, --visdom VISDOM
                        Host visdom server

-r, --root :- specifies the directory that houses the images and the annotations. (it is expected that the xml files are in root/annotations/ in PASCAL VOC format and the images in JPG format are in root/training/)

-e, --epoch :- If resuming training, epoch to start from (make sure epoch.pth exists in model_data/model and model_data/opt)

-l, --layers :- is the resnet version to use (either 50, 101 or 152)

-v, --visdom :- set this to true if you have visdom installed and would like to monitor your progress

Example:-

python3 train.py -r Data/ -e model_data/model/570.pth -l 101 -ne 100 -se 10 -v True

For predicting

run

python3 predict.py -h 
usage: predict.py [-h] [-r ROOT] [-s SPLIT] [-sc SCALE] [-sl SAVE_LOCATION]
                  [-t THRESHOLD]
                  model_name

Count rebars in images

positional arguments:
  model_name            The full path of the trained model (.pth file)

optional arguments:
  -h, --help            show this help message and exit
  -r ROOT, --root ROOT  Root directory housing the data
  -s SPLIT, --split SPLIT
                        Either test or custom split
  -sc SCALE, --scale SCALE
                        Scale down image by this factor
  -sl SAVE_LOCATION, --save_location SAVE_LOCATION
                        Location to save the predicted images
  -t THRESHOLD, --threshold THRESHOLD
                        Threshold to consider as positive, (predictions >
                        threshold) will be counted as 1

-r, --root :- root directory housing the data

-s, --split :- either test or a custom split (the directory should be structured as Data/test/ for test or Data/custom_name for a custom testing, which should contain all the images in jpg format to be tested)

-sc, --scale :- Use this argument to scale down the images. (sometimes larger images take a long time to predict with marginal improvements in accuracy)

-sl, --save_location :- Path to a directory to save the predictedi images

-t, --threshold :- consider predictions as positive if they are greater than threshold

Example:-

python3 predict.py -r Data/ -s testing/ -sc 5 -sl predicted_images/ -t 0.95

Acknowledgements

This project is done under the guidance of TATA innovation. The project is based on the paper https://arxiv.org/abs/1807.09856. If you would like to cite this work, please cite the original authors.

Citations

@Article{laradji2018blobs,
    title={Where are the Blobs: Counting by Localization with Point Supervision},
    author={Laradji, Issam H and Rostamzadeh, Negar and Pinheiro, Pedro O and Vazquez, David and Schmidt, Mark},
    journal = {ECCV},
    year = {2018}
}