Week 6-Warmth Of Image
Title: Weather Condition Prediction from Image
Team Members: Berk GÜLAY, Samet KALKAN, Mert SÜRÜCÜOĞLU
Emails Respectively: berkgulay.cs@gmail.com, abdulsametkalkan@gmail.com, mertsurucuogluu@gmail.com
In the previous week, we explained the progress of the project. And we show our methodology. Lastly, we show the collected results.
In this week, we have prepared a project presentation. We have already finished most of the work. Creating a dataset, a feature extracting and testing it with CNN Decision Tree, Random Forest, and SVM. And we explained everything about the project in the presentation.
At the end of the work, we saw that CNN and Random Forest Classifiers gave best overall and class-based accuracy scores. Decision Tree Classifier did not disgrace but slower than Random Forest. So was not worth to use. SVM did not fit our problem for real. It gave bumpy results, not fast, not cost efficient. Color Histogram, Brightness, Contrast, Sharpness and Haze metrics gave good performance since Random Forests gave also very well accuracy results using these features.
We obtained that, the cloudy class is tricky for our problem. Because every weather condition can have cloudy views as well. Our foresight is abstracting cloud class and integrating another class like thunder/storm would increase class based and overall accuracy far preponderant.
Our Best Results
For Random Forest we got, Cloudy~ %56, Sunny~ %76.5, Rainy~ %65, Snowy~ %60, Foggy~ %60, Overall~ %64
For CNN we got, Cloudy~ %45, Sunny~ %73.3, Rainy~ %61, Snowy~ %70, Foggy~ %70, Overall~ %65
We didn’t start to write the Final Report yet. Hopefully, in the next week, we will be writing our Final Report. We are really excited to see a last version of the work.
