This repository contains the code for the paper "Detecting Oncoming Vehicles at Night in Urban Scenarios - An Annotation Proof-Of-Concept".
The dataset is available here on kaggle. Please download the dataset before you continue with the next steps.
The dataset only contains the images and original bounding box annotations. To generate the binary masks, follow the next steps.
The data is stored scene-wise. This means that the data for each recorded scene is stored in a separate directory. Each scene directory then contains an image and several annotations folders. For example, this is the structure of scene 3:
00003
--- images
--- image1.jpg
--- image2.jpg
--- ....
--- imageN.jpg
--- annotations1
--- image1.json
--- image2.json
--- ....
--- imageN.json
--- annotations5
--- image1.json
--- image2.json
--- ....
--- imageN.json
--- annotations8
--- image1.json
--- image2.json
--- ....
--- imageN.json
The scene was annotated from annotators 1, 5, and 8, thus containing the three distinct annotation directories.
This repo requires the pytorch & torchvision nightly builds. To install, do the following:
First, install PyTorch nightly:
python3 -m venv venv
source venv bin activate
pip install numpy
# adjust cuda version if necessary
pip install --pre torch --extra-index-url https://download.pytorch.org/whl/nightly/cu117
Install torchvision:
git clone https://github.com/pytorch/vision.git
cd vision
git checkout d2d448c71b4cb054d160000a0f63eecad7867bdb
python setup.py develop
pip install flake8 typing mypy pytest pytest-mock scipy
Install this repo as a package:
cd ..
pip install -e .
If you want to train the models, login to Weights & Biases:
wandb login
python3 scripts/create_mask_dataset.py --param-path dataset/best_config.json --data-dir <path/to/dataset/> --mask-subdir majority-best-params --aggregation majority
python3 scripts/create_k_fold_dataset.py --data-dir <path/to/dataset/> --out-dir <path/to/5-fold-dataset/> --k 5
python training/train_kfold.py --data-dir <path/to/5-fold-dataset/> --num-epochs 200 --learning-rate 0.001 --gamma 4 --out-dir out/ --run-name 5fold-resnet50 --mask-subdir masks/majority-best-params --backbone resnet50 --no-empty-masks
This notebook walks you through reproducing all the numbers from the analysis of the annotation PoC in the paper.
The scripts to perform a grid search for all the four segmentation methods mentiond in the paper are located in scripts/grid_search/. For example:
cd scripts/grid_search/
python3 grid_search_adapt_gaussian.py --data-dir <path/to/dataset>
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent