ThorfinnYOLO: A hundred neural networks weekend

2026-07-19 · Source: Data Science on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Intermediate, medium

Summary

An extensive experimental campaign, dubbed "ThorfinnYOLO," involved training over a hundred variants of the YOLO26 model to detect a Bengal cat from an IP camera feed. The author conducted seven rounds (R0-R6) of experiments, totaling 97 runs, to test various hypotheses and optimize performance. Initial experiments (R0) established a baseline with `yolo26s` achieving mAP50 of 0.6844 at 640px. Subsequent rounds (R1-R3) explored hyperparameter sweeps, revealing that SGD with learning rates 0.03/0.05 collapsed training, 960px image size reduced mAP50, and a weight decay of 0.001 significantly degraded performance. The best R3 configuration (AdamW, lr 0.002, batch 16, mosaic 0.5) reached mAP50 0.8139. R4 indicated that larger YOLO26 models (n, s, m, l, x) offered no measurable gains on the small dataset. The pivotal R6 involved enriching the dataset from 83/16/5 to 357/67/22 images, resulting in a 54.6% relative gain in mAP50-95, reaching 0.6240. Notably, `YOLO26n` achieved the overall best mAP50 of 0.9044, demonstrating that dataset size was the dominant factor over hyperparameter tuning.

Key takeaway

For Machine Learning Engineers optimizing object detection models, prioritize dataset enrichment over extensive hyperparameter tuning. Your efforts in expanding and refining your training data will yield significantly greater performance gains than exhaustive sweeps of learning rates or augmentation settings. Consider starting with smaller models like YOLO26n, as they can achieve superior detection scores with less computational overhead, especially on expanded datasets. Be cautious with strong augmentation and weight decay, as these can degrade performance if not precisely matched to your specific dataset and optimizer.

Key insights

Dataset size and quality significantly outweigh hyperparameter tuning for object detection model performance.

Principles

• Dataset size is paramount for model quality.
• Smaller models can outperform larger ones.
• Augmentation strength must match dataset.

Method

The author conducted iterative rounds of YOLO26 training, starting with a baseline, sweeping hyperparameters (optimizer, LR, augmentation, weight decay, image size), and progressively enriching the dataset based on best-performing models.

In practice

• Prioritize dataset expansion over hyperparameter sweeps.
• Test smaller model variants like YOLO26n first.
• Carefully tune weight decay for adaptive optimizers.

Topics

• YOLO26
• Object Detection
• Dataset Augmentation
• Hyperparameter Tuning
• Computer Vision
• Model Optimization

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

Related on AIssential

• Breaking the Bottleneck: Achieving Native NMS-Free Inference with YOLO26 — This code provides a full pipeline for ONNX-based YOLOv8n object detection from image loading to result saving.
• Augmentation techniques for video surveillance in the visible and thermal spectral range — Augmentation techniques are crucial for robust multispectral CNN object detection, especially with limited thermal data.
• 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.
• FPN Paper Walkthrough: Leveraging the Internal Pyramid — FPN enriches multi-scale feature maps with both high spatial and semantic information for improved object detection.
• MoEIoU: Rethinking Bounding-Box Regression as a Mixture of Experts — MoEIoU adaptively prioritizes localization error components for improved bounding-box regression, leading to faster convergence and higher accuracy.
• Neural Network HyperParameters — Hyperparameters define neural network structure and training, requiring careful tuning for optimal performance and generalization.

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 Data Science on Medium.