A study of YOLOv11s for household waste detection: Baseline evaluation to model optimization

Exponential growth of improperly sorted household waste poses urgent environmental and economic challenges, particularly in Indonesia, where 50% of landfill waste originates from households. Contamination from mixed organic and inorganic materials drastically reduces recyclable value and complicates waste management. This study addresses these barriers by optimizing YOLOv11s for household waste detection through iterative dataset refinement and edge deployment strategies. Initial baseline evaluation on the TACO dataset achieved 21.2% mAP50 without augmentation, highlighting the complexity of real-world waste scenes. Progressive refinement of our custom dataset (11,876 instances) improved performance from 69.6 to 71.7% mAP50. Critically, static quantization reduced model size by 50% (18.3 MB to 9.7 MB) while achieving 72.6% mAP50, enabling deployment on resource-constrained edge devices. Results confirm YOLOv11s effectively handles complex visual characteristics of household waste—including high intra-class variability and contamination—without requiring extensive architectural modifications. This approach addresses critical barriers to effective recycling by enabling automated detection directly at waste sources with minimal infrastructure requirements.