KIT | KIT-Bibliothek | Impressum | Datenschutz

Uncertainty-guided dynamic routing for cost-efficient image classification

Rajoo, Edwin; Landgraf, Steven ORCID iD icon 1; Ulrich, Markus ORCID iD icon 1
1 Institut für Photogrammetrie und Fernerkundung (IPF), Karlsruher Institut für Technologie (KIT)

Abstract (englisch):

Deep learning networks have reached high accuracy in image classification, but often at the expense of high computational cost, since large models apply their full complexity uniformly to every input regardless of difficulty. We propose an uncertainty-guided dynamic routing framework pairing a lightweight network (MobileNetV3-Large) with a high-capacity network (EfficientNet-B3), routing each input individually based on uncertainty signals, entropy and confidence, from the lightweight model, together with its intermediate feature representation. The router is trained with a binary objective rewarding forwarding only when it corrects a mistake the small model would otherwise make, combined with a class-imbalance-aware weighting term and a regularizer discouraging unnecessary escalation. On Tiny ImageNet, the routed system reaches 78.54% accuracy, matching and, at the chosen operating point, slightly exceeding the standalone larger model (78.02%), while forwarding only around 31% of inputs to it – an estimated 57% compute reduction. An ablation study shows the two scalar uncertainty signals alone route as well as, or better than, the full signal set, with the 960-dimensional feature embedding contributing little once class imbalance is accounted for.


Verlagsausgabe §
DOI: 10.5445/IR/1000199235
Veröffentlicht am 17.11.2026
Cover der Publikation
Zugehörige Institution(en) am KIT Institut für Photogrammetrie und Fernerkundung (IPF)
Publikationstyp Proceedingsbeitrag
Publikationsjahr 2026
Sprache Englisch
Identifikator ISBN: 978-3-7315-1502-9
KITopen-ID: 1000199235
Erschienen in Forum Bildverarbeitung 2026 = Image Processing Forum 2026. Ed.: T. Längle
Veranstaltung Forum Bildverarbeitung (2026), Karlsruhe, Deutschland, 05.11.2026 – 06.11.2026
Verlag KIT Scientific Publishing
Seiten 91-102
Schlagwörter Dynamic routing, uncertainty estimation, adaptive, inference, image classification, efficient deep learning
Relationen in KITopen
KIT – Die Universität in der Helmholtz-Gemeinschaft
KITopen Landing Page