COWC Car Counting

This repository privides some python scripts and jupyter notebooks to train and evaluate convolutional neural networks which count cars from COWC aerial images.

Dependency

• NVIDIA Driver
• NVIDIA Docker

Usage

0. Clone this repo

$ PROJ_DIR=~/cowc_car_counting  # assuming you clone this repo to your home directory

$ git clone https://github.com/motokimura/cowc_car_counting.git $PROJ_DIR

1. Download COWC dataset

Download COWC aerial images and annotations from FTP server.

$ cd $PROJ_DIR/src/data
$ bash download_cowc.sh

2. Build Docker image

Build docker image to setup the environment to preprocess COWC dataset and train/evaluate convolutional neural networks.

$ cd $PROJ_DIR/docker
$ bash build.sh

In case you don't want to use docker, you have to install additional dependencies described in Dockerfile.

3. Preprocess COWC dataset

Generate crops from COWC train/val scenes defined in this text file.

Run docker container by following:

$ cd $PROJ_DIR/docker
$ bash run.sh

Now you should be inside the docker container you ran. Generate train/val crops by following:

$(docker) cd /workspace/src/features
$(docker) python gen_train_val_crops.py

In $PROJ_DIR/data/cowc_processed/train_val/crop/data, you should find many crops in which image and label are placed side by side like shown below. If you are interested in how train/val crops are sampled, please see this notebook.

Then, compute mean image over train crops. This will be used to normalize crops before input them to the network.

$(docker) cd /workspace/src/features
$(docker) python compute_mean.py

4. Train ResNet50

I reccomend you to use pretrained ResNet50 caffemodel to initialize model weights. It makes training much faster and the model more accurate.

Download ResNet-50-model.caffemodel from OneDrive download of this page. Then, move this file under $PROJ_DIR/models/caffemodels by following:

$ mkdir -p $PROJ_DIR/models/caffemodels
$ cd <directory in which ResNet-50-model.caffemodel is placed>
$ cp ResNet-50-model.caffemodel $PROJ_DIR/models/caffemodels

Train ResNet50 with generated crop images by following:

$(docker) cd /workspace/src/models
$(docker) python train_model.py --caffemodel ../../models/caffemodels/ResNet-50-model.caffemodel

If you want to train from scratch:

$(docker) cd /workspace/src/models
$(docker) python train_model.py

You can check training status and validation accuracy from TensorBoard:

# Open another terminal window outside the container and type:
$ cd $PROJ_DIR/docker
$ bash exec.sh

# Now you should be inside the container already running. Start TensorBoard by following:
$(docker) tensorboard --logdir /workspace/models

Then, open http://localhost:6006 from your browser.

5. Evaluate the network

Evaluate ResNet50 with jupyter notebook.

Launch jupyter notebook by flollowing:

$(docker) cd /workspace/notebooks
$(docker) jupyter notebook

Then, open http://localhost:8888 from your browser.

Note that you may need to modify the path to pre-trained model defined in the notebooks below.

5.1 Quantitative evaluation

Open this notebook to see quantitative evaluation on the test scenes defined in this text file.

5.2 Qualitative evaluation

Open this notebook to see qualitative evaluation on a test scene.

Distribusion of the cars counted by the network in a test scene (Salt Lake City):

5.3 Class activation mapping

Open this notebook to see class activation mapping on validation crops.

Output examples:

License

MIT License