Present repository is my workspace where I train and test models, and also develop script for deployement. It is not very clean though.
There is four jupyter notbooks with workflow of trainig models described below: Net_1.ipynb, Net_2.ipynb, Net_3.ipynb, Net_4.ipynb.
Also, there is notebook with workflow for training Yolo model: Yolo_train.ipynb
Folder MyUtils contains utils from Faster_RCNN team and a few of mine own.
Folder Metashape_scripts contains main script for working software detect_buldings_0_4.py and one I xperiment with during work. Also there is requirements.txt file for main script.
Data was collected from inner company source. It was photos taken from drone flying above villages. Data was annotated via Label Studio and Roboflow (compleat datasets was created via last)
Faster-RCNN architecture applyied for living buildings detection on imges taken from drone. Exmples shown below.
Keypoint-RCNN architecture applyied for buildings corners detection on images. Exaples shown below.
Work not finished. Idea was to detect power pillars and its ground point.
Mask-RCNN applyied for detection and segmentation of roads. Examples below.
Also, Yolov8 model was trained for the same task. Examples are shown in the following section
Python script was written for deployment models to specific working software. Scripts starts with dialog window with configurations possibilities as shown below
Results are as follows:
Note: Red dots show countours of roads detected and segmented via Mask-RCNN. Blue ones via Yolov8.
