Official repository of "Towards High-Quality Image Segmentation: Improving Topology Accuracy by Penalizing Neighbor Pixels (CVPR 26).
SCNP is an efficient method to improve topology accuracy by simply applying Min and MaxPooling to the logits in a particular way. We show its effectiveness on 13 datasets, including medical, non-medical datasets, datasets with tubular and non-tubular structures, single- and multi-class datasets, and on semantic and instance segmentation.
This repository contains our implementation of SCNP, the files required to run our experiments.
🚀 Project page: https://jmlipman.github.io/SCNP-SameClassNeighborPenalization/
📅 21-Feb-2026: SCNP has been accepted at CVPR 2026.
SCNP and example: here
Code to run our experiments: here