The author discusses their project in their Computer Vision class where they explore extracting frames from comic books. They highlight an algorithm that uses Canny to detect frames but has limitations with irregular polygons and negative frames. The author then describes creating a synthetic dataset of comic books and fine-tuning SAM to detect frame corners. They discuss the importance of canonicalizing corner point predictions and using L1 loss. Evaluation results show their method performing well on both procedurally generated and real-world comics compared to other methods. The author acknowledges shortcomings but appreciates the flexibility and potential of their approach. They provide access to their dataset, code, and invite feedback for improvement.
https://vrroom.github.io/blog/2024/02/23/comic-frame-segmentation.html