Eddie Park, PhD Candidate
Dec 16, 2022, 11:30am EC4-2101A
Images often have local redundant information that can strain a model in training or inference. An effective way to reduce spatial redundancy and complexity is to over-segment with superpixels. However, it is challenging to train a model with low spatial inductive bias. This talk will explore both the input representation of superpixels and the model that can effectively fit on highly variable graph-like structures.