VisDrone

OpenMMLab framework to train yolo's on VisDrone dataset currently only support Object Detection in Images.

Data

download this files.

trainset (1.44 GB): GoogleDrive
valset (0.07 GB): GoogleDrive
testset-dev (0.28 GB): GoogleDrive

Data stracture

please change the names for your split folders to: train, val, test. your data stracture need to be look like this:

root/
|_train
    |_images
    |_annotations
|_val
    |_images
    |_annotations
|_test
    |_images
    |_annotations

Quick Start

git clone https://github.com/eitan159/VisDrone.git
cd VisDrone
pip install -r requirements.txt

Data PreProcess

creating COCOForamt json files for VisDrone data.

NOTE: I modified the data, I only took samples that has the specific classes that I needed. See preprocess.py for better understanding :)

python VisDrone/preprocess.py --data_root your_path

params
--data_root: Path for your root dir that contains the data splits dirs

Run training

Please change the config files before training Run PP-YOLOE+:

python VisDrone/main.py --data_root your_path

params
--data_root: path for your data splits directories.
--work_dir (default exp/): dir name for all your saved ckpt, logs, etc.
--wandb_project (default VisDrone): wandb project name.
--model_size (default s): model version of PPYOLOE+ to train. Only supported with s/x
--lr (default 0.001)
--epochs (default 30)

eitan159 / VisDrone