Unfortunately, there is very little use case for my trained network in this algorithm. I've made a quick openCV algorithm to extract cards from the image, and it works decently well:Īt the moment, it's fairly limited - the entire card must be shown without obstruction nor cropping, otherwise it won't detect at all.
They're of course slightly worse than annonymous detection and impractical for any large number of cardbase, but it was an interesting approach. As you may imagine, this isn't sustainable for 10000+ different cards that exists in MTG, but I thought it would be reasonable for classifying 10 different cards. I tried to do an alternate approach - instead of making model identify cards as annonymous, train the model for EVERY single card. I can score about 20~25fps on my tiny YOLO, without using GPU. They have done quite an amazing job, and the speed isn't too bad, either. Thankfully, OpenCV had an implementation for DNN, which supports YOLO as well. I suppose there is a poor man's alternative - feed individual frames from the video into the detection script for image. Other darknet repos are in the same situation. There is a python example in the original repo of this fork, but it doesn't support video input. I tried to train it further, but it was already saturated, and was the best it could get.īad news, I couldn't find any repo that has python wrapper for darknet to pursue this project further. I've decided to continue with tiny YOLOv3 weights. Not to mention that it's quite slower, too.
UPCOMING MTG CARD SETS FULL
The training for full YOLOv3 model has turned sour - the loss saturated around 0.45, and didn't seem like it would improve in any reasonable amount of time.Īs expected, the performance of the model with 0.45 loss was fairly bad. At this rate, it should reach 50k epoch in about a week :/ Sept 13th, 2018 The author of darknet did mention that full network will take significantly more training effort, so I'll just have to wait. I've been training a new model with a full YOLOv3 configuration (previous one used Tiny YOLOv3), and it's been taking a lot more resources: The video demo can be found here: Sept 10th, 2018 I've ran a quick training with tiny_yolo configuration with new training data, and Voila! The model performs significantly better than the last iteration, even against some hard images with glaring & skew! The first prediction model can't detect anything from these new test images, so this is a huge improvement to the model :) Recompiled darknet with OpenCV and CUDNN installed, and recalculated anchors. Sept 7th, 2018Īdded several image augmentation techniques to apply to the training set: noise, dropout, light variation, and glaring:Ĭurrently trying to generate enough images to start model training. I'll have to think more about the first problem, though. The second and third problems should easily be solved by further augmenting the dataset with random lighting and image skew.
Fairly fragile against any glaring or light variations.Fails to spot some of the obscured cards, where only a fraction of them are shown.However, there are some blind spots on the model, notably: Uploading all the progresses on the model training for the last few days.įirst batch of model training is completed, where I used ~40,000 generated images of MTG cards laid out in one of the pre-defined pattern.Īfter 5000 training epochs, the model got 88% validation accuracy on the generated test set. setup_train.py: create train.txt and test.txt required to train YOLO from the training dataset.transform_data.py: generate training images using the aggregated card images and database.fetch_data.py: aggregates card images and database from.You can still find the files used to train them: Initially, the project used a powerful neural network named 'You Only Look Once (YOLO)' to detect individual cards, but it has been removed as of Oct 12th, 2018 (note) in favour of classical CV techniques.
0 kommentar(er)
