I attempted the Kaggle Airbus Ship Detection Challenge to show my level of compentancy with image segmantation
The goal of this challenge is to build a machine learning model to anaylze satelite images of container ships, located the ships and put a bounded box segment around them. By evaluating the F2 score at different intersection over union (IoU) threshold. The algorithm will sweep through a range of IoU and calculate the F2 score at each point(pixel). Each score is deteremined by the number of true positives(TP),false positives(FP) and false negativee(FN). A TP indicates a single preedicted object matches a ground truth object, FP shows a predicted object has no associated ground truth object and FN sa a ground truth object has no associated predictions. With this the average F2 score is calcualted.
The solution was implemented by using the U-net architecture. Which is a convolutional neural network that operates by downsampling and encoding the information in the image and than later upsampling and decoding the collected information from the same image.
- Python 3.8.16
- Tensorflow 2
- Jupyter Notebook
- U-net Model
- Dmitry Larko,- Kaggle Airbus Ship Detection Challenge Slides
- Jeff Heaton Deep Learning
- Cynical Learning Rate
- Albumentationns
- Kaggle Meetup Ship Detection Challenge
- Dmitry Larko, H2O.ai - Kaggle Airbus Ship Detection Challenge *2021, Installing TensorFlow 2.4, Keras, & Python 3.8 in Mac OSX Intel *Github Readme Cheatsheet