2026-10-09T15:48:16Z https://data.mendeley.com/oai

oai:data.mendeley.com/f4zn5p5d52.3 2026-09-22T17:05:04Z

wang, zhengbo Enhancement Without Replacement: Route-Preserving Residual Fusion for Low-Light Visible–Thermal Pedestrian Detection Mendeley Data This dataset contains the reproducibility materials supporting the manuscript "Enhancement Without Replacement: Route-Preserving Residual Fusion for Low-Light Visible–Thermal Pedestrian Detection" (submitted to Machine Vision and Applications). The deposited materials include model configurations for ConcatFusion-s, ConcatFusion-s-wide, RoutePreserve-s no-KD, residual-source variants, and insertion-position variants; split manifests for LLVIP, the FLIR person-only subset, and the KAIST night person-family subset; recorded training and evaluation arguments; processed source tables underlying the manuscript result tables (Tables 2–9); per-seed matched-initialization LLVIP records (seeds 0, 1, and 2); per-seed FLIR and KAIST results for seeds 0, 1, and 2; plotting scripts for Figures 3 and 4; software- and hardware-environment records; and available checkpoint metadata. Archived file names and verbatim command transcripts retain the earlier internal designation "AIRFuse" / "AIRFuse-s no-KD" for the RoutePreserve method; a full name mapping is provided in the archive README. The original LLVIP, FLIR, and KAIST images are not redistributed and must be obtained from their respective dataset providers under the applicable terms. Some legacy initialization artifacts are no longer available; the available provenance information and this limitation are documented in PROVENANCE_NOTES.md. Artificial Intelligence Computer Vision Signal Processing Image Processing Deng, Taorui Liu, Yukun wang, guangpeng Ma, Juncan cao, jing tang, haojun Dataset doi:10.17632/f4zn5p5d52.3 oai:data.mendeley.com/f4zn5p5d52.3 Creative Commons Attribution 4.0 International http://creativecommons.org/licenses/by/4.0 https://data.mendeley.com/datasets/f4zn5p5d52 2026-09-22T17:05:04Z