Skip to content

sahi benchmark on visdrone and xview datasets using fcos, vfnet and tood detectors

License

Notifications You must be signed in to change notification settings

JakeRobertBaker/small-object-detection-benchmark

 
 

Repository files navigation

My modifications

To train on yolox python mmdet_tools/train.py mmdet_configs/xview_yolox/yolox_crop_300_500_cls_60_nopipe.py

We have changed the requirments but are yet to update the requirments.txt file. We ran

conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=11.1 -c pytorch -c conda-forge
pip install sahi>=0.9.3
pip install pillow
pip install mmdet==2.21.0
pip install mmcv-full==1.4.3 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.10.0/index.html
pip install tensorboard>=2.7.0
pip install scipy
pip install setuptools==59.5.0

We have also created files to clean and restrict classes to only the trucks and cars,

python xview/xview_to_coco_cars_trucks.py --train_images_dir="data/xview/train_images" --train_geojson_path="/vol/bitbucket/jrb21/project/xView/data/xView_train.geojson" --output_dir="data/xview/coco" --is_val=True --clean=True

The last two boolean flags determine whether we are using the validation set or if we are 'cleaining' aka removing negative coordinates which exist in the xview dataset.

About

sahi benchmark on visdrone and xview datasets using fcos, vfnet and tood detectors

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Jupyter Notebook 99.5%
  • Other 0.5%