Semantic segmentation of ariel images taken from drone
- Developed a deep learning based model for segmentation of pixels in images taken from the drone into 12 categories.
- Used Deeplab-v3 architecture along with Resnet-101 as a backbone.
- Used fine-tuning on model pre-trained on Imagenet.
- Used various data augmentation techniques as the number of images in the dataset was small.
- Used focal loss as a loss function to overcome the class imbalance.
- Achieved dice score of 0.796 on the test set.
- Tools used: Pytorch, OpenCV, Albumentations, Matplotlib, Kaggle Kernels
Kaggle competition link: https://www.kaggle.com/c/opencv-pytorch-course-segmentation
Dataset download link: https://www.kaggle.com/c/opencv-pytorch-course-segmentation/data