ladd-and-weights

Metadata of Lacmus dataset and trained model weitghts tracked with dvc.

Usage

Repository in intended to sustain reproducability and versioning of the dataset and model weights.

• Install our fork of dvc, becouse original dvc has an issue

conda create -n lacmus-dvc python==3.9 pip
conda activate lacmus-dvc
git clone https://github.com/lacmus-foundation/dvc.git
cd dvc
pip install .

• In order to get access to cloud storage, contact lacmus team as described on wiki and get your credentials file.
• Put credentials to ~/.aws/
• Configure dvc and get data:

git clone https://github.com/lacmus-foundation/ladd-and-weights.git
cd ladd-and-weights
dvc remote modify --local digital_ocean credentialpath ~/.aws/credentials
dvc pull

• to gather LADD from relevant parts run python merge.py to merge with heridal dataset and convert to yolov5 format for training refer to https://github.com/lacmus-foundation/lacmus-research/blob/master/extra/EDA/LADDvsIPSAR_merge_and_prepare.ipynb

Contributing

In case you'll add new imagesSet, please also update dvc.yaml and gather_LADD.sh

Project structure

\
- dataset
	-- pretrain
		--- sdd-lacmus-version - Prepared Standford Dron Dataset
		--- VisDrone2019-DET VisDone Dataset
	-- LADD
		---
		...  folders part of dataset, gathered as project evolves 
		--- 
	-- unmarked
		---
		... folders of file sets, submitted by users, but not yet included into main dataset
		---
	-- 3rd_party
		--- heridal - http://ipsar.fesb.unist.hr/HERIDAL%20database.html (*) 
- weights 
	-- yolo5 - wieghts for yolo v5 (https://github.com/ultralytics/yolov5) currently in production 
	-- keras-retinanet - weights for keras retina-net model (https://github.com/lacmus-foundation/lacmus/tree/master/keras_retinanet) - previous version of network
	-- torch
		--- pretrain
		--- experimental

---

(*) Dunja Božić-Štulić, Željko Marušić, Sven Gotovac: Deep Learning Approach on Aerial Imagery in Supporting Land Search and Rescue Missions, International Journal of Computer Vision, 2019.