Aerial-Semantic-Segmentation-Drone-using-Pytorch

Aerial Semantic Segmentation of Drone Captured Images, using Pytorch.

Project Repository: https://github.com/mr-ravin/Aerial-Semantic-Segmentation-Drone-using-Pytorch

Dataset Related Information

• Aerial Image Dataset: https://www.kaggle.com/datasets/santurini/semantic-segmentation-drone-dataset

• Number of Raw Classes: 24, indexed from 0 to 23. After processing this data, these classes are grouped into 5 classes, indexed from 0 to 4.

index	Class
0	unlabeled
1	paved-area
2	dirt
3	grass
4	gravel
5	water
6	rocks
7	pool
8	vegetation
9	roof
10	wall
11	window
12	door
13	fence
14	fence-pole
15	person
16	dog
17	car
18	bicycle
19	tree
20	bald-tree
21	ar-marker
22	obstacle
23	conflicting

Final Number of Grouped Classes: 5

grouped class id	individual classes	group color in rgb	grouped class name
0	0, 6, 10, 11, 12, 13, 14, 21, 22, 23	[155,38,182]	obstacles
1	5, 7	[14,135,204]	water
2	2, 3, 8, 19, 20	[124,252,0]	nature
3	15, 16, 17, 18	[255,20,147]	moving
4	1, 4, 9	[169,169,169]	landable

Visualise a sample from the dataset

Directory Structure

|-- dataset/
|      |-- images/ # contains all rgb images as .jpg
|      |-- masks/  # contains all ground truth masks as .png
|      |-- train/
|      |     |-- images/
|      |-- val/
|           |-- images/
|-- utils
|    |-- preprocess.py
|    |-- compute.py
|    |-- grephics.py
|
|-- train_files.txt
|-- valid_files.txt
|-- dataloader.py
|-- run.ipynb
|-- weights/       # it will contain the weight file after model training.
|-- results/
       |-- view_dataset.jpg
       |-- inference_results.jpg
       |-- overall_analysis.jpg

Information regarding the images present in our train set is present inside train_files.txtand image information for validation set is present inside valid_files.txt. It's provided so that one can reproduce the results.

Steps to Train the model

• Download the dataset and place its images and masks inside dataset/images/ and dataset/masks/ respectively.
• Run the jupyter notebook: run.ipynb

Model Performance

We have trained the model upto 25 epochs and have shared the performance matrices and some sample results below.