Machine Learning Primer
A Thorough Guide to SegFormer — Why a Transformer Without Positional Encoding Works for Segmentation
This article breaks down SegFormer, introduced by NVIDIA, The University of Hong Kong, and collaborators at NeurIPS 2021, based on the original paper. We examine the hierarchical MiT encoder without positional encoding, the parameter-efficient All-MLP decoder, the effective receptive field, and the measured mIoU, parameter counts, and compute from B0 to B5 with equations.