Federico Carrara

SWITi: Quantifying and Reducing Tiling Artifacts with Sliding Window Inner Tiling

Federico Carrara, Aman Kukde, Melisande Croft, Joran Deschamps, Florian Jug

SWITi is a test-time method for reducing artifacts in tiled predictions, particularly for neural networks that learn posterior distributions from which solutions are sampled at inference time. Tiled predictions are unavoidable for large image data, and artifacts arise whenever ti…Read more