TRANSFER LEARNING: PASSENGERS AND HELMET USE ON MOTORCYCLES IN MYANMAR AND NEPAL
Mobility can differ among different places. Researchers have studied how people move in Myanmar and Nepal, focusing on the usage of motorcycles. Much effort has been put into identifying, in an automated way, motorcycles from video footage using deep learning techniques and classifying the position of riders and helmets. This project has applied transfer learning from a labelled traffic dataset from Myanmar to classify the position encoding of passengers and helmets on motorcycles to a dataset from Nepal, with the aim of utilising the datasets similarities to construct a model for the latter one. The results are promising, showing the model trained on Myanmar has performed even better on the Nepal dataset.
├── app.conf <- Shared settings across the repo
├── README.md <- The top-level README for developers using this project
├── notebooks <- All notebooks
| ├── inspect_model.csv <- Inspect the achitecture of the save model
| ├── train_val_acc_loss_plots.ipynb
│ └── within_class_graphs.ipynb
├── src
| ├── bike_detection.py <- Bike detection using YOLOv3
│ ├── classification_encode.py <- Classification Encoding
│ ├── data_split.py <- Data Splitting
│ ├── merge_annotations.py <- Merge annotation files
│ └── utils.py <- Utility functions
├── scripts
| ├── cp_nepal.sh <- Copy Nepal data set from shared folder
| ├── dl_myanmar.sh <- Download Myanmar data set from the web
│ └── run_on_hpc.sh <- Run classification_encode.py on HPC with GPU
│
└── requirements.txt <- The requirements file for reproducing the analysis environment
Find below all the instructions to set up your enviroment, run Jupyter Lab and use DTU HPC with GPU.
virtualenv env
source env/bin/activate # MacOS
.\env\Scripts\activate # Windows
pip install --upgrade pip
pip -r requirements.txt
# cd into project folder
source env/bin/activate # MacOS
.\env\Scripts\activate # Windows
jupyter lab
git clone https://github.com/eriklindernoren/PyTorch-YOLOv3.git pytorchyolo
bash ./pytorchyolo/weights/download_weights.sh
mv -t pytorchyolo/weights/ yolov3-tiny.weights yolov3.weights darknet53.conv.74
# Then run:
python src/bike_detections.py
The model can be trained running python src/classification_encode.py.
The script requires some arguments to be executed:
usage: classification_encode.py [-h] [--load LOAD] [--train | --no-train] [--tuning | --no-tuning]
[--save SAVE]
project
positional arguments:
project Project to run.
optional arguments:
-h, --help show this help message and exit
--load LOAD (Optional) Saved model to load.
--train, --no-train (Optional) if model needs to be trained.
--tuning, --no-tuning
(Optional) if model needs to be fine tuned.
--save SAVE (Optional) Filename for saving the model.
Be sure to adapt the bash script accordingly for the desired output.
# Make sure a logs file exists
mkdir logs
# Adapt the bash script according to the model you want to run, then run
bsub < ./scripts/run_on_hpc.sh
Extra storage has been requested to DTU for this project. Since the provided data for Myanmar wasn't completed, we created our own copy. Moreover, to facilitate the preprocessing we made a copy of the Nepal dataset as well.
bash ./scripts/dl_myanmar.sh
python src/merge_annotations.py Myanmar # check
python src/data_split.py Myanmar
bash ./scripts/cp_nepal.sh
python src/merge_annotations.py Nepal
python src/data_split.py Nepal
- Christina Nørgaard Bartozzi
- Erla Hrafnkelsdóttir
- Pietro Rampazzo
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent