What is YOLOv6? A Deep Insight into the Object Detection Model

Athulya Sundaresan Geetha Department of Computer Science, Huddersfield University, Queensgate, Huddersfield HD1 3DH, UK; * Correspondence: [email protected]; · 2024

Read What is YOLOv6? A Deep Insight into the Object Detection Model by Athulya Sundaresan Geetha Department of Computer Science, Huddersfield University, Queensgate, Huddersfield HD1 3DH, UK; * Correspondence: [email protected]; (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-13006.txt.

Introduction

Athulya Sundaresan Geetha gives a deep tour of YOLOv6, a real-time object detection model. The YOLO family spots objects in one fast pass. Version six pairs an EfficientRep backbone for feature pull with a Rep-PAN neck that fuses features across scales. Her 2024 review walks the design, its training tricks, and its scores on the COCO benchmark. Small builds trade a little accuracy for blazing speed.

The nano build tops a thousand frames per second on a Tesla T4 card. Larger builds beat same-class rivals on accuracy while holding real-time pace. The biggest variant claims cutting-edge accuracy for live use. The piece serves engineers picking a detector for cameras, drones, and factory lines.

Catalog record

CREATORAthulya Sundaresan Geetha Department of Computer Science, Huddersfield University, Queensgate, Huddersfield HD1 3DH, UK; * Correspondence: [email protected];
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-13006

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

What is YOLOv6? A Deep Insight into the Object Detection Model Athulya Sundaresan Geetha Department of Computer Science, Huddersfield University, Queensgate, Huddersfield HD1 3DH, UK; * Correspondence: [email protected]; Abstract This work explores the YOLOv6 object detection model in depth, concentrating on its design framework, optimization techniques, and detection capabilities. YOLOv6’s core elements consist of the EfficientRep Backbone for robust feature extraction and the Rep-PAN Neck for seamless feature aggregation, ensuring high-performance object detection. Evaluated on the COCO dataset, YOLOv6-N achieves 37.5% AP at 1187 FPS on an NVIDIA Tesla T4 GPU. YOLOv6-S reaches 45.0% AP at 484 FPS, outperforming models like PPYOLOE-S, YOLOv5-S, YOLOX-S, and YOLOv8-S in the same class. Moreover, YOLOv6-M and YOLOv6-L also show better accuracy (50.0% and 52.8%) while maintaining comparable inference speeds to other detectors. With an upgraded backbone and neck structure, YOLOv6-L6 delivers cutting-edge accuracy in real-time. August 24, 2026 K eywords Object Detection; YOLO; Convolutional Neural Networks; Real-Time Image processing; Computer Vision; YOLOv6 Architecture 1 Introduction Computer vision is a fast-moving field that provides machines [ 1 , 2 ] with the ability to process and comprehend visual signals. Object detection is a crucial element of this field, involving the accurate localization and identification of objects within images or video streams [ 3 ] . Features were extracted from images both horizontally and vertically using the Histogram of Oriented Gradients (HOG) and Vertical Histogram of Oriented Gradients (VHOG) techniques [ 4 ] . Support Vector Machine and Extreme Learning Machine were then used for classification. The SVM method relied on visualizing features by localizing images and features during dynamic motion in spatiotemporal areas for classification but faced challenges in accurately detecting movements [ 5 , 6 , 7 , 8 ] . The inefficiencies of two-stage approaches, such as their extended runtime, dependence on advanced architectures, and need for constant manual intervention, led to the rise of CNN models for classification and detection. When used together, Faster R-CNN and GoogleNet achieved better results in object detection than the previous models [ 9 ] . PoseNet, MaskR-CNN, and MobileNet were used in another study to refine image segmentation and boost performance. Other models, such as Region-based Convolutional Network method (R-CNN) [ 10 ] , Google-Net [ 11 ] , Fast-RCNN [ 12 ] , AlexNet [ 13 ] , ResNet [ 14 ] , Faster-RCNN [ 15 ] , and VGG-Net [ 16 ] , were used to improve object identification further [ 17 ] . Even with performance improvements, challenges such as efficient data management, real-time execution, and more advanced architectures persist. YOLO (You Only Look Once) models address these limitations, overcoming the drawbacks of two-stage detection methods. YOLOv1, based on DarkNet, includes 24 convolutional layers, and Fast YOLO includes 9 layers [ 18 ] . YOLOv2 is a refined version built on DarkNet-19, providing batch standardization, box bounding optimization, in-depth class identification, intricate feature extraction [ 19 ] , and size-oriented dimension [ 20 ] . YOLOv3, based on the DarkNet-53 architecture, employs residual networks for efficient feature extraction and binary cross-entropy for training optimization [ 21 ] . YOLOv4, built on YOLOv3’s predictive head, features CSPDarkNet53 and integrates PANet with SPP as the backbone and neck for enhanced performance, respectively [ 22 ] . The high-performance speed of YOLOv5 is achieved through the combination of the enhanced CSP-PAN neck and the SPPF head [ 23 ] . Not only does YOLOv6 offer enhanced performance speed, but it also reduces the complexity of computational processes [ 24 ] . With the addition of an auxiliary head and a lead head, YOLOv7 achieves superior object detection accuracy [ 25 ] . In YOLOv8, a modified CSPDarkNet53 backbone and PAN-FPN neck enable object identification, image segmentation, and movement tracking. The inclusion of Programmable Gradient Information and Generalized Efficient Layer Aggregation Network in YOLOv9 and YOLOv10 enhances detection accuracy, with YOLOv10 also removing non-maximum suppression [ 26 , 27 ] . The goal of this study is to evaluate the YOLOv6 object detection model in relation to other YOLO models, focusing on accuracy and inference speed across its various variants (n, s, m, l, x). In addition to performance metrics, the research aims to analyze architectural differences, training methodologies, and computational efficiency to provide a comprehensive comparison. The study also investigates YOLOv6’s adaptability and its scalability for varying levels of complexity, offering insights into its practical usability in real-time environment. 2 Evolution of YOLO 2.1 YOLO and YOLOv2 YOLO revolutionized object detection with its speed and efficiency but faced limitations. Due to its grid-based structure, YOLO had trouble detecting smaller objects in clusters, struggled with shapes, and suffered from localization inaccuracies caused by its loss function treating small and large box errors equally [ 28 , 29 ] . YOLOv2 employed the Darknet-19 architecture. Batch normalization enhances performance by 2% mAP and avoids overfitting through regularization. A higher input resolution of 448×448 increases accuracy by 4% mAP for superior performance. Anchor boxes are used instead of connected layers to enhance recall. K-means clustering optimized the precision-recall balance via anchor box selection [ 30 ] . 2.2 YOLOv3 YOLOv3 introduces Darknet-53, a new 106-layer network architecture featuring residual blocks and upsampling networks. This design is larger, faster, and more accurate than the previous Darknet-19 used in YOLOv2. Key advancements include improved bounding box predictions through a logistic regression model, enabling more precise objectness scoring. For class predictions, YOLOv3 replaces the softmax approach with independent logistic classifiers, allowing better handling of complex scenarios with overlapping labels. Additionally, YOLOv3 enhances prediction accuracy by performing three-scale predictions for each location in the input image, capturing fine-grained semantic details for superior output quality [ 31 ] . 2.3 YOLOv4 YOLOv4 achieves a remarkable balance of speed and accuracy in object detection, outperforming YOLOv3 with a 10% improvement in accuracy and a 12% boost in speed. With 29 convolutional layers and 3 × 3 filters, CSPDarknet53 has an estimated 27.6 million parameters. Key enhancements over YOLOv3 include the addition of a Spatial Pyramid Pooling block, which expands the receptive field and isolates important contextual features without slowing down the network. YOLOv4 replaces YOLOv3’s Feature Pyramid Network (FPN) with PANet to improve parameter aggregation across detection layers. It introduces advanced data augmentation techniques, such as the mosaic method. Hyper-parameters are optimized using genetic algorithms, further boosting the model’s performance and robustness [ 32 ] . 2.4 YOLOR You Only Look One Representation (YOLOR) [ 33 ] is a unified network for multiple tasks, combining explicit (conscious) and implicit (experiential) knowledge to enhance performance. By aligning features, optimizing predictions, and standardizing representations for multi-task learning, this integration enhances the architecture. Precision improves by roughly 0.5% in YOLOR by embedding implicit representation in each FPN’s feature map, with further refinement of predictions coming from its inclusion in the network’s output layers. For multi-task learning, YOLOR incorporates canonical representation to mitigate the performance degradation often caused by joint optimization of loss functions across tasks, achieving better performance than other models. 2.5 YOLOX The backbone of YOLOX [ 34 ] is Darknet-53, integrated into an updated YOLOv3 structure. The decoupled head in YOLOX splits the processes of detecting objects and determining their spatial coordinates. YOLOX also incorporates

Read the complete text