Data for study "Evaluating the Impact of Gaussian Noise on YOLO-Based Object Detection: A Systematic Study on Noise Robustness and Denoising-Retraining Strategies"

---------------------------Basic information---------------------------1. Journal article: Evaluating the Impact of Gaussian Noise on YOLO-Based Object Detection: A Systematic Study on Noise Robustness and Denoising-Retraining Strategies 2. DOI: 10.5281/zenodo.20842880 3. Contact information Name: Vo Thanh Kiet Institution: VSB – Technical University of Ostrava E-mail: [e-mail] ORCID: https://orcid.org/0009-0002-3278-8755 4. Dataset publication date: 2026-06-25 5. Place of publication: Ostrava, Czechia ------------------------------------------------------------------6. Dataset Description================================================================================ This dataset contains the experimental results and the source code used to producethe results and figures presented in the article, which evaluates the robustness ofYOLO-based object detectors under Gaussian noise and the effectiveness of imagedenoising combined with model retraining. Source image dataset (PnPLO). The study uses the PnPLO (Person and Person-LikeObjects) human-detection dataset, a publicly available dataset distributed on Kaggleunder the slug "karthika95/pedestrian-detection"(https://www.kaggle.com/datasets/karthika95/pedestrian-detection). PnPLO contains1,339 images annotated in PASCAL VOC format (XML) with two object classes: "person"(real people) and "person-like" (human-like hard negatives such as statues,mannequins, robots and posters), and includes weather-augmented composite samples.The images have variable native resolution (from 181 × 188 up to 4128 × 2752 pixels,most around 500 × 375) and are resized to 640 × 640 pixels at training time. Theimages were downloaded programmatically through the kagglehub API and the VOCannotations were converted to YOLO format inside the pipeline. Experimental design. Five YOLO variants (YOLOv8m–YOLOv12m) were trained on PnPLO(split 70% train / 15% validation / 15% test) and tested under Gaussian noise atσ ∈ {0, 1, 5, 10, 20, 30}, where σ = 0 denotes the clean baseline. Six denoisingconfigurations — a no-denoising baseline, Gaussian Filter, BM3D, DnCNN, Autoencoder,and CAE+PSO — were applied prior to retraining, following a unified three-stageprotocol: (i) clean training, (ii) noisy inference, and (iii) denoised retraining andre-evaluation. Detection performance (mAP@50, mAP@50–95, Precision, Recall) and aperson vs person-like classification analysis (Accuracy, F1-Score, Sensitivity,Specificity) were recorded for every model × noise level × denoiser combination. Note on the image data: PnPLO is a third-party dataset and its original images andannotations are NOT redistributed in this package; they remain publicly availablefrom the original provider at the Kaggle link above. This package provides thequantitative results and the code required to reproduce the study from the originalimages. The source code is also maintained at:https://github.com/kietvt3195/YOLO-Based-Object-Detection When using the source code, please follow the instructions provided in the notebookheader. IMPORTANT: Do not modify the file locations within the folder structure. Folder Structure-------------------------------------------------------------------------------- yolo_noise_robustness_dataset (contains 2 subfolders and 1 readme file)│├── results (detection and classification metric tables in .csv)│ ││ ├── all_results.csv│ └── classification_metrics.csv│├── source_code│ └── YOLO_Denoise_Experiment_Karthy_COLAB_NUMBERED_Final_Ver_4.ipynb│└── readme.txt File Descriptions-------------------------------------------------------------------------------- • results/all_results.csv Object-detection metrics for every combination of YOLO model (YOLOv8m–YOLOv12m) × noise level (σ ∈ {0, 1, 5, 10, 20, 30}) × denoising method (no-denoising baseline, gaussian_filter, bm3d, dncnn, autoencoder, cae_pso). Columns: model, noise_sigma, denoise_method, mAP50, mAP50-95, Precision, Recall, Composite. Here σ = 0 is the clean baseline. This file contains the mAP-based results that underlie the headline performance figures of the article (e.g. mAP@50 = 0.8992 for YOLOv8m on clean data). • results/classification_metrics.csv Person vs person-like classification metrics derived from the same experimental runs, expressed as a two-class confusion-matrix analysis. Columns: model, noise_sigma, denoise_method, Accuracy, F1-Score, Sensitivity, Specificity, Precision, TP, FP, FN, TN. • source_code/YOLO_Denoise_Experiment_Karthy_COLAB_NUMBERED_Final_Ver_4.ipynb Jupyter notebook implementing the complete pipeline — Kaggle download and VOC→YOLO conversion, Gaussian-noise injection, the six denoising configurations, YOLO training and evaluation, metric aggregation, and figure generation. The cells are numbered to follow the three-stage protocol. Mirror of the public GitHub repository.-------------------------------------------------------------------------------- 7. Funding: This work was supported in part by the European Regional Development Fund under the project Research Platform for Digital Transformation and Society 5.0 CZ.02.01.01/00/23_021/0012599 within the Jan Amos Komensky Operational Program. This work was supported in part by the Ministry of Education of the Czech Republic (Project No. SP2026/012, SP2026/075).