研究

Ultralytics 出版物

了解我们团队在各个领域的研究和出版物。

Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models

Presents YOLO26, a unified real-time vision model family with native NMS-free end-to-end inference, no DFL, MuSGD training, Progressive Loss, and STAL label assignment. Across five scales and multiple vision tasks, YOLO26 advances the COCO accuracy-latency Pareto front and adds YOLOE-26 open-vocabulary inference.

arXiv (2026) · Glenn Jocher, Jing Qiu, Mengyu Liu, Shuai Lyu, Fatih Cagatay Akyon, Muhammet Esat Kalfaoglu

出版年份期刊概览
2026IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops
SenBen: sensitive scene graphs for explainable content moderation
Fatih Cagatay Akyon, Alptekin Temizel
Introducing SenBen: a Visual Genome-style scene graph benchmark for explainable sensitive content moderation (13,999 frames, 25 objects, 28 attributes, 14 predicates, 16 tags). And a compact 241M vision-language model trained with vocabulary-aware recall (VAR) loss fits in 1.2 GB VRAM, runs 7-23x faster than open-weight VLMs, and outperforms every baseline except the Gemini 3 teacher.
2022IEEE ICIP
Slicing aided hyper inference and fine-tuning for small object detection
Fatih Cagatay Akyon, Sinan Onur Altinuc, Alptekin Temizel
Proposed SAHI framework: a generic slicing-aided inference and fine-tuning pipeline for small object detection on top of any pretrained detector. Up to +6.8 AP (inference) and +14.5 AP (fine-tuning) on Visdrone and xView. Supports Ultralytics framework models.
2022IEEE
Exudate regeneration for automated exudate detection in retinal fundus images
Muhammad Hussain, Hussain Al-Aqrabi, Muhammad Munawar, Richard Hill, Simon Parkinson
This paper presents a framework for automated exudate detection in retinal images, introducing a CES mechanism using Ultralytics YOLOv5 for data generation and a lightweight CNN for classification, achieving accuracy with efficient deployment via synthetic data and self-labeling.
2022Computer Science Conference Proceedings (CSCP)
Ultralytics YOLOv5, YOLO-X, YOLO-R, YOLOV7 performance comparison: a survey
Ismat Saira Gillani, Muhammad Rizwan Munawar, Muhammad Talha
This conference paper explores the evolution of the YOLO model series, tracing its progress from Ultralytics YOLOv5 to YOLOv7, highlighting key advancements in speed and accuracy using pre-trained models on the COCO dataset.
2021IEEE AVSS
Track boosting and synthetic data aided drone detection
Fatih Cagatay Akyon, Ogulcan Eryuksel, Kamil Anil Ozfuttu, Sinan Onur Altinuc
Drone detection pipeline (1st place at the AVSS 2021 Drone vs. Bird Challenge): a Ultralytics YOLOv5 detector fine-tuned with real plus synthetic data, augmented by a Kalman-based tracker that boosts confidence under weak contrast, long-range, and low-visibility conditions.
2016NIM A
Multi-photon disambiguation on stripling-anode multi-channel plates
G.R. Jocher, M.J. Wetstein, B. Adams, K. Nishimura, S.M. Usman
Presents a maximum a posteriori technique for resolving near-simultaneous multi-photon events on stripline-anode Large-Area Picosecond Photo-Detectors (LAPPDs), improving timing and position accuracy in high-rate particle detection scenarios.
2015Scientific Reports
AGM2015: Antineutrino global map 2015
S.M. Usman, G.R. Jocher, S.T. Dye, W.F. McDonough, J.G. Learned
A comprehensive experimentally informed global model of Earth's antineutrino flux across the 0–11 MeV spectrum, providing a reference map for reactor monitoring, neutrino mass hierarchy studies, and detection of undeclared nuclear facilities.
2014Proceedings of SPIE
Minimum separation vector mapping (MSVM)
G.R. Jocher, J.T. Dolloff, P.J. Doucette, B.M. Hottel, H.J. Theiss
Introduces MSVM, an algorithm for extracting 3D structure from GPS-tagged 2D video that natively exploits GPS metadata, consistently outperforming traditional Structure from Motion approaches in aerial ISR imagery.
2014Proceedings of SPIE
Optimal full motion video registration with rigorous error propagation
J. Dolloff, B. Hottel, P. Doucette, H. Theiss, G. Jocher
Describes a mathematically rigorous approach to registering full motion video frames to geospatial reference data, including formal error propagation for use in precision ISR and situational awareness applications.
2013Physics Reports
Theoretical antineutrino detection, direction and ranging at long distances
G.R. Jocher, D.A. Bondy, B.M. Dobbs, S.T. Dye, J.A. Georges III, J.G. Learned, C.L. Mulliss, S. Usman
Introduces NUDAR (NeUtrino Direction and Ranging), a theoretical framework for passively detecting, directionally localizing, and ranging nuclear reactors using antineutrino emissions over long distances, with implications for non-proliferation monitoring.

让我们携手构建 AI 的未来！

开启你的旅程，迎接机器学习的未来

获取 YOLO 许可开始使用

Ultralytics 出版物

SenBen: sensitive scene graphs for explainable content moderation

Slicing aided hyper inference and fine-tuning for small object detection

Exudate regeneration for automated exudate detection in retinal fundus images

Ultralytics YOLOv5, YOLO-X, YOLO-R, YOLOV7 performance comparison: a survey

Track boosting and synthetic data aided drone detection

Multi-photon disambiguation on stripling-anode multi-channel plates

AGM2015: Antineutrino global map 2015

Minimum separation vector mapping (MSVM)

Optimal full motion video registration with rigorous error propagation

Theoretical antineutrino detection, direction and ranging at long distances

让我们携手构建 AI 的未来！