jsiloto / context_change_bdd100k

Context Change for Split Computing in Autonomous Driving

Current State

This repository currently is about object detection on BDD100k repository

Environment

Download this repo, install the requirements, download the YOLOV5 fork

git clone [e-mail]:jsiloto/context_change_bdd100k.git
cd context_change_bdd100k
virtualenv venv
source venv/bin/activate
pip install -r requirements.txt
git clone [e-mail]:jsiloto/yolov5.git

Dataset

Download the dataset to $DATASET

DATASET=/data/datasets/bdd100k/
ls -l $DATASET
> bdd100k_images_100k.zip  bdd100k_labels_release.zip 
unzip $DATASET/bdd100k_images_100k.zip bdd100k/images/* -d $DATASET/..
unzip -j $DATASET/bdd100k_labels_release.zip bdd100k/labels/* -d $DATASET/labels_bdd100k

Prepare Annotations

Annotations need to be prepared on the YOLO formate Link Here For that we use the approach of converting first to the coco format (BDD100k->COCO)

python scripts/bdd2coco.py '--bdd_dir' $DATASET

Now we convert the coco format annotations to the yolo format (COCO->YOLO)

python scripts/coco2yolo.py '--bdd_dir' $DATASET

Other Dataset Specs

Domain specific datasets

generate the domain specific datasets:

python scripts/create_domain_datasets.py '--bdd_dir' $DATASET

Movie Dataset

Create the frame by frame dataset

First be sure to download the appropriate files for reproduction and unzip to the correct location

DATASET=/data/datasets/bdd100kvideo/
ls -l $DATASET
> bdd100k_videos_train_14.zip  bdd100k_videos_val_04.zip
unzip -j $DATASET/bdd100k_videos_train_14.zip bdd100k/videos/* -d $DATASET/video/train
unzip -j $DATASET/bdd100k_videos_val_04.zip bdd100k/videos/* -d $DATASET/video/val

Run the following scripts which will create the frame by frame dataset, annotate it using the yolov5 format and evaluate on domain specific models

./scripts/movie_dataset_build.sh
./scripts/movie_dataset_domain.sh

Now, we have built and evaluated the movie dataset with many model, we can see which models are better together

python scripts/domain_switching_map.py

export CUDA_VISIBLE_DEVICES=1
python yolov5/train.py --img 640 --epochs 3 --data bdd100k.yaml --weights yolov5s.pt