Extremal Contours: Gradient-driven contours for compact visual attribution

A new visual attribution method for vision models has been introduced, emphasizing the creation of smooth, tunable contours rather than traditional dense masks. This approach aims to provide faithful and compact explanations, addressing the need for more interpretable visual attributions without relying on extensive post-processing. The technique utilizes a truncated Fourier series to represent contours and optimizes them using classifier gradients. By focusing on gradient-driven contours, the method offers a novel way to generate clear and concise visual explanations. This development represents a shift from dense mask-based attributions toward more compact and interpretable visualizations. The approach is designed to enhance the clarity and usability of visual explanations in machine learning models. Overall, it contributes to the ongoing effort to improve interpretability in vision systems.