# Context Beginner set of 16,000 custom images for categorizing polyhedral dice # Content ## Image Organization ~ 85% / 15% (train / valid) - all training images are 480x480 - all d4, d8, d10, and d12 validation images are 480x480 - most d6 and d20 validation images are 480x480 - a small percentage of additional d6 and d20 validation images are larger (1024px long side) and completely unlike the training set ## Methodology _All images were created, edited, and sorted by Mario Lurig._ - Fixed camera positions (minimum 2 angles) used to capture video on a [rotating platform][1] with two white lights - Minimum 5 different dice used on 6 different backgrounds (white and various colors) - Video was then exported as images and then batch cropped to 480x480 - Handheld camera moved over 5+ dice on various wood surfaces (minimum 2) using natural lighting - Video edited and exported to images then batch cropped to 480x480 - Images that were partially out of crop were manually removed The additional d6 and d20 validation images were from my personal image collection or taken additionally on a variety of surfaces with no care for lighting conditions to work as a more robust test. The validation images were pulled from the full image set (480x480 images) as a 1/7th slice rather than randomly. If preferred, you could combine train/valid together and randomly assign them via your code; this data organization method was chosen to help beginners. Finally, images taken in like groups are named in like ways. For instance, d4_angleXXX are all d4 dice taken at an angle. d10_top are all d10 dice taken from the top down. Once again done in an effort to make it easy to add/remove data and see how that changes the results. _**Note:** There are more d6 and d20 images than d4,d8,d10,d12 due to those two dice being my initial test set before building the rest._ Inspiration As an avid roleplayer and the person behind [HeroMuster.com](https://heromuster.com/), I decided to start learning ML from not only the code and execution side, but also from the data collection/organization side. This felt like a great way to do that. Quicky results [ResNet101 0.9947 accuracy][2] ![Confusion Matrix][3] [1]: https://www.thingiverse.com/make:502149 [2]: https://www.kaggle.com/ucffool/fast-ai-resnet101-99-5-or-better [3]: https://i.imgur.com/QXY9zJ2.png