ICE-Z and Multi-Point SAGFiLM: Illumination-Conditioned Dynamic Feature Modulation for YOLO26 in Adverse-Lighting Detection

2026-99-2436

11/20/2026

Authors
Abstract
Content
Under severe illumination shifts, the detection representation space is disrupted, ultimately manifesting as weakened object boundaries, foreground-background activation imbalance, and abnormal response diffusion during multi-scale feature fusion. To address this issue, we propose a targeted enhancement framework for YOLO26 under illumination-hard conditions. The framework preserves the original backbone-neck-head architecture and performs conditional enhancement only on key intermediate features during the forward pass of the DetectionModel. Specifically, a 5-dimensional global illumination condition vector is extracted from the deep SPPF feature via ICE-Z; three independent FiLMHeads then predict dedicated modulation parameters for the three critical nodes, namely the high-level semantic node A, and the multi-scale pyramid nodes P3 and P4; finally, SAGFiLM performs condition-driven dynamic feature correction with residual reinjection.
Compared with the YOLO26x baseline of the same model scale, the final model improves Precision, Recall, mAP50, and mAP50-95 on BDD100K-hard from 65.69%, 49.02%, 57.51%, and 35.66% to 75.18%, 51.00%, 59.33%, and 36.26%, respectively. On VisDrone-hard, the corresponding metrics increase from 90.66%, 85.43%, 91.84%, and 64.42% to 91.38%, 86.35%, 92.65%, and 64.98%, respectively. Meanwhile, the method introduces only about a 0.31% increase in parameters and an additional training overhead of approximately 8.71%-10.36%. We also report its performance degradation on the full-set distribution, indicating that the proposed method is a robustness enhancement mechanism tailored to a specific working regime rather than a universal plug-in applicable across arbitrary distributions.
Meta TagsDetails
DOI
https://doi.org/10.4271/2026-99-2436
Citation
Li, Z. and Li, Y., "ICE-Z and Multi-Point SAGFiLM: Illumination-Conditioned Dynamic Feature Modulation for YOLO26 in Adverse-Lighting Detection," 2026 2nd International Conference on Intelligent Transportation and Future Mobility (ITFM2026), Xi'An, China, March 27, 2026, https://doi.org/10.4271/2026-99-2436.
Additional Details
Publisher
Published
11 hours ago
Product Code
2026-99-2436
Content Type
Technical Paper
Language
English