Answers are grounded in this paper and cite their sources.
Explanations released by a real-time object detector are often validated at full precision and assumed to remain stable after edge quantization. We test this assumption for a frozen YOLOv10x detector with a fused Grad-CAM teacher and a distilled forward-only surrogate under FP16 and INT8 deployment in five regimes from BDD100K and VisDrone-VID2019. Detection retention is measured by relative mAP and explanation retention by agreement between deployed and full-precision surrogate maps. Under INT8, explanation agreement fell by 5.8–11.4%, 5.4–7.3 times the noise floor of a 1/255 input perturbation, while mAP fell by 1.7–7.1%; on pairs not used to develop the surrogate, the loss was 9.6%. Controls attributed about three quarters of the loss to the feature-change step of a fixed-box-first decomposition, and showed that the asymmetry persisted on training-excluded sequences and across training seeds and calibration sets and that the surrogate’s maps depend on the learned weights. Two regimes repeated on a Jetson AGX Orin gave similar results. A motion-aware transport operator gave model-estimated reuse ages from 26 frames in daylight to six in dense traffic. Accuracy alone therefore does not establish that a detector’s explanations are retained after deployment.
No takes yet. Share an insight, caveat, or question.
Bahri et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: