Will it Blend? Synthetic Datasets with Blender - Decoding Chessboards | Part 3 скачать в хорошем качестве

Will it Blend? Synthetic Datasets with Blender - Decoding Chessboards | Part 3

Hey folks, We make a high-quality synthetic dataset of chessboards and chess pieces with Blender and train a YOLO-based object detection model with it. Today's model is trained to find chessboards in every-day images and videos, but we're especially setting the stage to create future custom models that could do that and more. Leave a comment if you enjoyed this video or want to learn certain parts more in-depth! Part 2 of the series:    • Can AI + Synthetic Data Beat Classic Compu...   Series:    • Teaching AI to See   00:00 Introduction 00:17 Part 1/2 Recap 01:08 Adding Chess Pieces and Segmentation 01:35 Trying Roboflow and Hand Labeling Challenges 02:34 Making Better Synthetic Data 03:27 Modeling Chessboards 04:17 Modeling Pieces 05:23 Procedural Piece Placement from PGNs 06:08 Generating Darknet/COCO Annotations 06:52 Segmentation Convex Hull Polygons and Masks 07:32 Randomizing EVERYTHING and Rendering 08:03 Background Image Augmentation 09:28 Training Chessboards 10:10 Training Pieces Too! 10:30 Testing Model on Videos 11:08 Adding Computer Vision Results! 12:05 Next Steps This video goes into what makes a machine learning dataset good or bad especially when it comes to bounding box and segmentation labeling for object detection, explaining how you might create this yourself. I also tried out the free tier of Roboflow and it was pretty useful to validate my dataset format, but we quickly switch back to local training. In Blender I show how I model several chessboards and chess pieces, and procedurally lay the pieces on the board using real game positions from chess matches, same for the randomized camera and lighting. Now that we can generate thousands of annotated images (including bounding boxes and segmentation mask polygons), we re-train our previous YOLO chessboard detection model, and also for pieces too, something we couldn't do before. With the new dataset, the model is able to detect chessboards and pieces in real-world images and videos much better than, despite being trained solely on synthetic data. I can't wait to use this dataset for more advanced machine learning approaches like U-net segmentations and corner detections. The dataset is linked here [WIP].