Due April 23, 2025 by 11:59pm
Points 150
In this part you should have your complete solution up and running, evaluated on the validation subset. Not yet perfect, but it should be ready for final testing. Usually, if you already implemented preprocessing, segmentation and feature extraction, the classification is left for this part. Again, customized projects may need discussion with Walter to agree on concrete deliverables.
What to do and deliver?
- A report (no page limit, but try to be concise; 1000-2000 words should suffice) as a separate "Part 4" section of the readme.md in your GitHub repository that includes:
- A justification of the choice of classifier. For instance, if you selected SVM with RBF kernel, say why you think this classifier is good for your project. (30 points)
- A classification accuracy achieved on the training and validation subsets. That is, how many objects were classified correctly and how many of them were classified incorrectly (It is better to provide a percentage instead of numbers). Students working on object detection may report Intersection over Union averaged over the training and testing samples. More advanced students can select the performance metrics that best suit their given problem (for instance, Precision-Recall, f-measure, plot ROC curves, etc.) and justify the use of these metrics. (20 points)
- A short commentary related to the observed accuracy and ideas for improvements. For instance, if you see almost perfect accuracy on the training set, and much worse on the validation set, what does that mean? If this is not desired, what do you think you could do to improve the generalization capabilities of your solution? You can think about one really small improvement (out of all improvements you propose in this report) to be implemented before the final testing. (60 points).
- For teams: explain individual contributions of each team member (this is needed to have this assignment graded).
- Push your current code to your project repository (40 points). This code should implement what you described in the report. Provide instructions how to run your code on a data example (attach this example to the code). Either Walter or Louisa will run them to see how the current solution works.