Publication
Efficient Vision Models for Jetson: Steel Classification via Knowledge Distillation
Yasin Esfandiari; Karan Rajshekar; Sabine Janzen; Wolfgang Maaß
In: Proceedings of the 1st Workshop on Sustainability and Resource-Efficiency of Artificial Intelligence (SuRE 2026) at IJCAI 2026. Workshop on Sustainability and Resource-Efficiency of Artificial Intelligence (SuRE-2026), located at IJCAI-ECAI 2026, August 17, Bremen, Germany, CEUR Workshop Proceedings, CEUR-WS.org, 2026.
Abstract
Deploying vision models at the industrial edge requires balancing accuracy against energy and latency constraints that server-oriented models cannot meet, yet prior distilled-ViT work measures efficiency through proxy metrics (FLOPs and server-GPU throughput) rather than actual on-device energy. We close this gap on the DOES nonalloyed steel-scrap benchmark (8,131 test tiles, eight EU steel grades), distilling ViT-Large and Vision-LSTM (ViL-Base, an xLSTM-based backbone) teachers into 5.5M-parameter DeiT-Tiny students and benchmarking every model on an NVIDIA Jetson Xavier NX with onboard power monitoring and latency measurement. Our distilled students achieve 88.81% and 87.07% top-1 with 2.93–3.56% accuracy gaps to their teachers, reducing inference energy by 15.9–34× and latency by 18.1–44×. To the best of our knowledge, this is also the first cross-inductivebias comparison in knowledge distillation for vision involving a recurrent xLSTM-based teacher: same-family distillation (ViT→DeiT, 88.81%) achieves 1.74 pp higher accuracy than cross-family (ViL→DeiT, 87.07%), yet both deliver equivalent on-device efficiency.