Method Details


Details for method 'NeuroFuser'

 

Method overview

name NeuroFuser
challenge pixel-level semantic labeling
details NeuroFuser is a memory-efficient encoder–decoder that frames multi-scale fusion as a budgeted, controllable operation rather than a fixed architectural design. It combines lightweight plug-in attention–fusion modules with a neuromodulation controller that regulates multi-head, multi-scale fusion, including windowing and dilation mixtures, to keep memory usage predictable while preserving fine spatial details across resolutions and improving robustness under domain shift without requiring backbone-specific modifications.
publication NeuroFuser: Resource-aware neuromodulation of multi-scale fusion attention for domain adaptive segmentation
Serdar Erişen, André Borrmann
Pattern Recognition
https://doi.org/10.1016/j.patcog.2026.114167
project page / code https://github.com/serdarch/NeuroFixer
used Cityscapes data fine annotations
used external data ImageNet
runtime 0.3 s
A100-80GB GPU
subsampling no
submission date May, 2025
previous submissions

 

Average results

Metric Value
IoU Classes 86.3824
iIoU Classes 73.8648
IoU Categories 93.0498
iIoU Categories 84.9874

 

Class results

Class IoU iIoU
road 98.9022 -
sidewalk 89.1463 -
building 94.9477 -
wall 73.03 -
fence 70.5467 -
pole 75.4043 -
traffic light 80.859 -
traffic sign 84.6852 -
vegetation 94.5162 -
terrain 75.4898 -
sky 96.2866 -
person 90.0917 78.7607
rider 79.9928 66.8587
car 96.7655 90.7805
truck 85.9444 67.4733
bus 95.8194 76.9838
train 94.3543 70.2545
motorcycle 82.2275 66.4652
bicycle 82.2557 73.3419

 

Category results

Category IoU iIoU
flat 98.8717 -
nature 94.2863 -
object 80.4353 -
sky 96.2866 -
construction 94.9693 -
human 90.0363 79.7282
vehicle 96.463 90.2467

 

Links

Download results as .csv file

Benchmark page