Year: 2026
Type: article
Abstract: Objects that closely match background color are difficult to localize because chromatic cues weaken and boundaries can be obscured. This study evaluates a YOLO11 baseline using a dataset containing at least 12 object categories and approximately 630 images per category, trained for 100 epochs. Training and validation curves are summarized alongside confidence-based and precision-recall analyses, and an exploratory grayscale-Canny-YOLO fusion path is documented separately from the evaluated baseline. The evaluation plot reports an overall [email protected] of 0.607; category AP ranges from 0.433 for Table to 0.831 for Bus, while the F1-confidence curve peaks at 0.60 near a confidence threshold of 0.284. The category-count plot indicates substantial imbalance among annotated instances. These results establish a measurable baseline and expose class-dependent performance variation. The findings support dataset balancing, fixed evaluation partitions, and ablation testing as necessary next steps for reliable same-color-background detection.
DOI: 10.54097/6kn0w081
Authors: Jie Xiong, Bo Huang, Chengguo Xiao, Wangjun Zou
Institution: Chengdu University of Information Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7233738920
