Repository for improved yolov5 for small target detection

DotaToYolo converts DOTA annotation format to YOLO annotation format in txt

VisDronetoYOLO converts VisDrone annotation format to YOLO annotation format in txt

pngtojpg converts png files to jpg in a folder

Repository also contains two jupyter notebooks: DataExplorationClasswise: This notebook contains all classwise pixel area size analysis of objects in both DOTA and VisDrone dataset DataExploration: this notebook calculates average pixel area of all the objects of all classes in both DOTA and VisDrone dataset

The trained weights can be downloaded from gdrive below:

Yolov5x 40 epochs DOTA

Yolov5x 40 epochs VisDrone

Yolov5 improved 40 epochs VisDrone

Yolov5 improved 40 epochs DOTA

These trained weights can be used with Yolov5 official github repository here to reproduce the results, the custom model file is also provided as yolov5-small-target.yaml that can be used in the official yolov5 models folder Alternatively, the version of repository used in this work can be downloaded from here