A Controlled Benchmark of Active Learning Strategies for Class-Imbalanced Waste Detection on SortWaste with YOLO
Year: 2026
Type: article
Abstract: Bounding-box annotation constrains waste detection under class imbalance and video redundancy. This benchmark compares Random, Plug and Play Active Learning (PPAL), Rarity-Aware, and an entropy/diversity strategy (Entropy) on SortWaste with YOLOv8n over six paired seeds and 5%–20% labeled-image budgets. The primary metric is normalized area under the budget curve (nAUBC) of mean average precision (mAP@50:95) across intersection-over-union (IoU) thresholds of 0.50–0.95. Mean nAUBC ± sample standard deviation was 0.2831 ±0.0079 for PPAL, 0.2780 ±0.0063 for Entropy, 0.2705 ± 0.0057 for Random, and 0.2661 ± 0.0057 for Rarity-Aware. At 20%, mean test mAP@50:95 was 0.3000, 0.2967, 0.2917, and 0.2817, respectively. The paired 20% differences from Random had intervals that included zero for PPAL, Entropy, and Rarity-Aware. Mean paired nAUBC differences from Random were+0.0126 for PPAL (unadjusted 95% paired t-interval [0.0032, 0.0220]) and +0.0075 for Entropy ([0.0020, 0.0129]). Each seed–budget point was trained once, and the intervals are not adjusted for multiple comparisons. PETOil contributed 46% and 53% of the PPAL and Entropy nAUBC gains. Active strategies selected images with more boxes and greater temporal concentration; their mean gains were smaller under a box-count budget.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7225004413