AICUP2022 - droneDetection

About the project

無人機飛行載具之智慧計數競賽 https://tbrain.trendmicro.com.tw/Competitions/Details/25

---

Getting Start

Installation

# clone the repository
git clone https://github.com/jimmylin0979/AICUP2022-droneDetection-v2.git

# go to repository folder
cd AICUP2022-droneDetection-v2

# pip install required packages
pip install -r requirements.txt

---

Dataset

We package the complicated setup procedure into a bash script, you can simply complete setup by following steps below.
To check what was happening during setting, please have a see in ./data/dataset.sh

1. Download dataset zip file and place them into below format:

   ./data
   ├── dataset.sh
   ├── Private Testing Dataset_v2.zip
   ├── Public Testing Dataset_v2.zip
   ├── Training Dataset_v5.zip
   ├── VisDrone2019-DET-train.zip
   ├── ...
   

2. Run the bash script to setup the dataset

   cd data
   bash dataset.sh

It will take about 20 ~ 30 minutes, depends on your CPU power, to complete the setup. (It took 20 minutes for i5-12400 to finish setup)
The final directory structure should be like (we only list some important dataset folders here):

./data
├── FusionDataset
│   ├── images
|   └── labels
|
├── Private Testing Dataset_v2
│   ├── private
|   └── private_tiled
|
├── Public Testing Dataset_v2
│   ├── public
|   └── public_tiled
|
├── train
│   └── public
│       ├── img1001.png
|       ├── img1001.txt
|       ├── ....
|       ├── img1500.png
│       └── img1500.txt
|
├── Training Dataset_v5
│   ├── images
|   └── labels
|
├── Training Dataset_v5 tiled
│   ├── images
|   └── labels
|
├── ...

---

Training

Be sure to start training after you finish all transformation on dataset (ex. tiling, gamma correction, and so on)

1. Download the pretrained weights of yolov7-e6e_training.pt in transfer learning section from https://github.com/WongKinYiu/yolov7, place the weights into folder yolov7/weights

2. Run the command below to start training, the results will be stored in ./results folder,
   the follow command will consume about 20G memory usage on GPU, you can modify the batch size depends on your personal device.

   cd yolov7
   python train_aux.py --workers 4 --device 0 --batch-size 4 --data data/fusionDataset.yaml --img 1280 1280 --cfg cfg/training/yolov7-e6e.yaml --weights weights/yolov7-e6e_training.pt --name yolov7-e6e-aug-tile-fusion --hyp data/hyp.scratch.custom.yaml --label-smoothing 0.1

---

Evaluating

Be sure to start training after you finish all transformation on dataset (ex. tiling, gamma correction, and so on)

• Weights (Private score 0.758381): https://drive.google.com/uc?export=download&id=1zGFK57FCeo-ylCeEoirI0ilJg1aFtAdE

1. Run the below command to start evluating, the predictions will be stored in ./results/yolov7/train/

   cd yolov7
   
   # Public dataset (tiling)
   python detect.py --weights ../results/yolov7/train/yolov7-e6e-aug-tile-fusion/weights/best.pt --source ../data/Public\ Testing\ Dataset_v2/public_tiled/data/ --img-size 1280 --conf-thres 0.4 --device 0 --save-txt --save-conf --nosave --augment --name yolov7-e6e-aug-tile-fusion-public
   
   # Private dataset (tiling)
   python detect.py --weights ../results/yolov7/train/yolov7-e6e-aug-tile-fusion/weights/best.pt --source ../data/Private\ Testing\ Dataset_v2/private_tiled/data/ --img-size 1280 --conf-thres 0.4 --device 0 --save-txt --save-conf --nosave --augment --name yolov7-e6e-aug-tile-fusion-private

2. Merge tiling predictions

   cd TiledSet
   
   # Public dataset (merged)
   python main.py --mode merge --root-src ../data/'Public Testing Dataset_v2'/public --root-dst ../results/yolov7/detect/yolov7-e6e-aug-tile-fusion-public/predictionstile.csv
   mv predictions.csv predictions_public.csv
   mv predictions_public.csv ../results/yolov7/detect/yolov7-e6e-aug-tile-fusion-public/predictions_public.csv
   
   # Private dataset (merged)
   python main.py --mode merge --root-src ../data/'Private Testing Dataset_v2'/private --root-dst ../results/yolov7/detect/yolov7-e6e-aug-tile-fusion-private/predictions_tile.csv
   mv predictions.csv predictions_private.csv
   mv predictions_private.csv ../results/yolov7/detect/yolov7-e6e-aug-tile-fusion-private/predictions_private.csv

3. Filter out confidence column

   cd utils
   
   # Public dataset
   python helper_filter.py --file ../results/yolov7/detect/yolov7-e6e-aug-tile-fusion-public/predictions_public.csv
   mv predictions.csv ../results/yolov7/detect/yolov7-e6e-aug-tile-fusion-public/predictions_public.csv   # the file now is able to upload to leaderboard now, 
   
   # Private dataset
   python helper_filter.py --file ../results/yolov7/detect/yolov7-e6e-aug-tile-fusion-private/predictions_private.csv
   mv predictions.csv ../results/yolov7/detect/yolov7-e6e-aug-tile-fusion-private/predictions_private.csv  # the file now is able to upload to leaderboard now,

4. The final prediction file will be stored in ./results/yolov7/detect/yolov7-e6e-aug-tile-fusion-private/predictions_private.csv and results/yolov7/detect/yolov7-e6e-aug-tile-fusion-public/predictions_public.csv separately

---

Visualization

# Public visualization
python main_visualization.py

# Private visualization
python main_visualization_private.py

Acknowledgments

• mmdetection
• yolov7
• Best-README-Template