YOLOv26 for Automated Inventory Management: A Study on Dense Object Detection in Retail Environments
Authors/Creators
- 1. College of New Energy, Hubei University of Automotive Technology, China
- 2. School of Automotive Intelligent Manufacturing, Hubei University of Automotive Technology, Shiyan, China
Description
Abstract:
For retail to manage its inventory well, it needs to be able to accurately detect objects. However, shelves that are full of products make this difficult because they can block each other, overlap, and change size. This study is the first to look at YOLOv26, a new YOLO architecture that doesn't need NMS for inference, uses Progressive Loss Balancing, and the MuSGD optimizer. It does this on the SKU-110K dataset, which has over 11,000 images and 1.7 million annotated objects. A small YOLOv26n model trained for 50 epochs on an NVIDIA RTX 3060 with the best hyperparameters (416×416 input, batch size 8, AdamW) gets 81.9% mAP50 and 47.6% mAP50-95, with 84.8% precision and 73.4% recall. Inference runs at 158 frames per second, which means it can be used in real time on low-end hardware. Analysis shows that 77% of objects are correctly detected. Most of the errors come from false positives in the background and items that are heavily blocked. YOLOv26n is 22% better than YOLOv8 in mAP50 and gets close to bigger models with very little extra processing power. Integrating with ByteTrack for counting inventory shows that it can be used in real life, setting a strong standard for efficient, edge-capable retail automation.
Keywords: YOLOv26, detecting objects, SKU-110K, managing inventory, crowded retail scenes, and detecting things in real time.