Background Shortcuts Drive High Accuracy in SAR Vehicle Recognition
November 18, 2026
Analysis of the MSTAR and ATRNet-STAR benchmarks reveals that deep networks rely heavily on background context rather than the vehicles themselves. Removing the vehicle from MSTAR chips drops accuracy from 99% to 73.3%, while the scene-change split on ATRNet-STAR drops accuracy to 22.8%.
HOW THIS AFFECTS YOU
●
researcherAvoid over-relying on benchmark scores for SAR models that may just be memorizing environmental textures.