| Title |
Analysis of Train?Validation Dependence and Evaluation Reliabilityin a Video-Derived Object Detection Dataset |
| DOI |
https://doi.org/10.5573/ieie.2026.63.10.124 |
| Keywords |
Object detection; Evaluation reliability; Train?validation dependence; Video-derived dataset; YOLO11 |
| Abstract |
Object-detection evaluation assumes sufficient independence between training and validation data, yet similar frames may cross nominal partitions in video-derived datasets. This study operationalizes train?validation visual dependence and audits the public Medical Pills dataset comprising 115 images and 2,022 annotated objects. Exhaustive analysis of all 6,555 image pairs identified long-range visual recurrence. In the official split, mean cross-split GT-box-region correlation was 0.0913, while the median nearest-training correlation of validation images was 0.8262. A performance-independent Minimum-dependence split reduced these values to 0.0427 and 0.2714, respectively. An initial three-seed YOLO11n comparison yielded an Official?Minimum mAP@0.5:0.95 difference of +0.0086, but the validation sets differed in composition. We therefore conducted a performance-blind five-fold paired experiment (50 runs) using identical validation images within each fold and evaluated pooled out-of-fold performance over all 115 images. The high-dependence condition yielded higher mAP@0.5:0.95 in 4 of 5 paired seeds, but the mean difference was only +0.00118 (95% CI: ?0.00230 to +0.00466), so the initial magnitude was not reproduced. We therefore make no definitive or causal performance claim. These results show that auditing split structure and sample-level dependence, rather than assuming independence from nominal partitioning alone, is important for improving the reliability of validation-performance interpretation in frame-based object-detection datasets. |