Synthetic and Derived Training Images for Campus Waste Detection: A Multi-Seed Evaluation with YOLOv8n

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Internet of Things (IoT) & Connected Devices, Data Science & Analytics · Depth: Expert, extended

Summary

A study evaluated the effectiveness of synthetic and derived images in improving a YOLOv8n detector for campus waste detection, using a real dataset of 148 photographs (86 for training, 31 for validation, 31 for testing). Researchers tested twelve joint-training configurations, repeating seven principal settings with four matched seeds. The real-only model achieved a mean mAP@0.5 of 0.691 [0.665, 0.722]. Adding background-replaced images reduced performance to 0.560 [0.499, 0.619], while isolated-object images yielded 0.680 [0.644, 0.724]. The full augmentation pool resulted in 0.487 [0.438, 0.537]. A corrected experiment with hand-and-forearm composites showed a paired difference of +0.034 [−0.063, 0.199], but no reliable benefit. None of the synthetic or derived image configurations surpassed the real-only baseline. Single-seed transfer experiments revealed source-dependent rankings. FP32 models (12.3 MB) processed 2.8–3.0 images/second on a four-thread CPU proxy, but an INT8 export (3.4 MB) failed to produce detections.

Key takeaway

For Machine Learning Engineers developing object detectors for edge devices with limited real-world data, you should prioritize collecting new, diverse real object instances over generating more transformations of existing objects. Your synthetic data must introduce deployment-relevant variation and be rigorously evaluated against a strong real-only baseline, using multi-seed runs to account for training noise. Also, diagnose INT8 export failures on target hardware before relying on quantization for deployment.

Key insights

Synthetic and derived training data did not improve YOLOv8n waste detection over real-only data, highlighting domain mismatch challenges.

Principles

• Image source matters more than quantity.
• Small datasets amplify training noise.
• Contextual cues can be critical.

Method

YOLOv8n was trained on a 148-photo campus waste dataset, augmented with synthetic/derived images (background replacement, isolated objects, hand composites). Performance was evaluated using mAP@0.5 across four matched seeds.

In practice

• Prioritize real deployment-view data.
• Evaluate synthetic data against strong baselines.
• Document image provenance carefully.

Topics

• Waste Detection
• Object Detection
• YOLOv8n
• Synthetic Data
• Data Augmentation
• Edge AI
• Transfer Learning

Code references

• ultralytics/ultralytics

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Synthetic and Derived Training Images for Campus Waste Detection: A Multi-Seed Evaluation with YOLOv8n — Synthetic and derived images did not improve YOLOv8n waste detection performance over real-only data in this study.
• YOLOv26: An Object Detector Built for Real-Time Deployment — YOLOv26 prioritizes edge deployment through architectural simplification and hardware-friendly optimizations.
• Training-Free Metrics for Synthetic Object Detection Data: A Proxy for Detector Performance — CCDM metrics provide a training-free, highly accurate proxy for evaluating synthetic object detection data utility, outperforming existing methods.
• Visual Prompting Meets Feature Reconstruction-Based Anomaly Detection with Dual-Teacher Supervision — Robust anomaly detection in varied real-world conditions is achieved by combining visual prompting, adaptive teacher models, and synthetic data.
• A.R.I.S.: Automated Recycling Identification System for E-Waste Classification Using Deep Learning — A.R.I.S. uses YOLOx for real-time e-waste classification, significantly improving material recovery efficiency.
• ThorfinnYOLO: A hundred neural networks weekend — Dataset size and quality significantly outweigh hyperparameter tuning for object detection model performance.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.