Code, trained weights and measurement outputs for: A Two-Stage Cascade Framework for Traffic Sign Detection - Reproduction and Boundary Analysis
Year: 2026
Type: software
Abstract: This archive accompanies the revised manuscript submitted to Scientific Reports (submission ID 514dfe79-1080-4b36-96ee-25c83e4af852). It contains the code, trained weights and measurement outputs required to regenerate every number, table and figure in the paper. What is implemented. The reproduction covers the cascade skeleton only: CLAHE + Gaussian + Wiener preprocessing, a YOLOv8n screening stage, crop-based re-check with the same detector, and a rule-based semantic gate (numeric legality and class exclusivity). What is not implemented. Cross-layer weighted fusion, the spatial-attention module, the ResNet-50 classifier and the BERT perplexity term are not part of this reproduction. No results or ablations for these components are claimed in the manuscript; this is declared in Section 3.1 and Section 5. What is not included. The raw images of TT100K and CCTSDB are third-party public datasets and are not redistributed. The README gives the official sources and the exact expected file counts so that the conversion scripts can be re-run. Contents. scripts/ - dataset preparation, training, the cascade pipeline, nine analysis scripts and two figure scripts runs/ - three YOLOv8n screening-stage checkpoints (seeds 0-2) and one RT-DETR checkpoint (7 epochs), in the directory layout the scripts expect results/ - the training log, ablation summary, condition-stratified results, module-level timings, threshold sweep, semantic-gate flip statistics, queueing results, RT-DETR evaluation, and the raw predictions for 3,016 frames that feed all analyses configs/ - the 45-class TT100K subset definition and the weather/size stratification of the CCTSDB evaluation images Caveats. The RT-DETR checkpoint was trained for 7 epochs only and reaches [email protected] = 0.0101 with recall 0.045, which is insufficient to serve as the second stage; the re-check stage in all reported results is crop-based re-detection with the same YOLOv8n. All latencies were measured on a consumer laptop with an NVIDIA MX450 and no real-time operation is claimed. The overload results come from a FIFO queueing simulation driven by measured service times, not from a live multi-stream test. See README.md for the step-by-step reproduction commands mapped to each paper item.
Institution: Xi'an Peihua University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225083221