YOLO
YOLO (You Only Look Once) is a series of deep learning-based object detection models originally proposed by Joseph Redmon et al. Its defining characteristic is the reformulation of object detection as a single-pass regression problem: a unified neural network simultaneously predicts object categories and bounding box coordinates in one forward pass. Compared with traditional two-stage detection methods, YOLO achieves faster inference speeds and is suited for real-time object recognition and localization in images and video.
Core Facts
The YOLO (You Only Look Once) object detection algorithm was first proposed by Joseph Redmon in 2015.
Since YOLO was proposed by Joseph Redmon in 2015, its core innovation has been to reformulate object detection as a regression problem solved by a single forward pass.
YOLO treats object detection as a single regression problem, simultaneously predicting bounding boxes and class probabilities in a single forward pass.
Timeline (last 90 days)
YOLO models degrade sharply in performance when deployed across datasets, mainly because the annotation conventions of the two datasets differ (fixed-size bounding boxes vs. bounding boxes that fit the lesion contour).
The updated content of this PhD's 13th large AI class includes the latest version of YOLO, the overall large model framework, and more.
Running inference first with a pretrained model (such as the YOLO series or Faster R-CNN) can cover 60%-80% of common objects.
Human detection is usually based on lightweight object detection models (such as the YOLO series or vendors' in-house models), running on the camera or on a local NVR.
RICO and VINS are both screenshot-domain data (pixel-level clear, no perspective distortion), and there is an obvious domain gap with the stroke noise and unclosed-line characteristics of hand-drawn wireframes, so when transferring directly it is recommended to use only their pretrained backbone networks and then fine-tune with a small number of hand-drawn samples.
Annotating 200-500 hand-drawn wireframes with LabelImg or CVAT, combined with data augmentation (rotation, scaling, stroke perturbation), is enough to obtain a demo-level YOLO model that can be shown.
At a sample size of 200-500, YOLO-series models (such as YOLOv8n and YOLOv8s) can usually converge to usable accuracy.
Lightweight deep learning models such as YOLO can perform real-time object detection on edge devices.
Faster R-CNN is a two-stage detector, with high accuracy but slow speed; YOLO unifies detection into a regression problem, making it extremely fast; the difference between the two is essentially the trade-off between "find first, then judge" and "one step to the end," with the YOLO series being more popular in real-time scenarios, while two-stage methods still hold a place on accuracy-sensitive academic benchmarks.
Mainstream 2D object detection methods can be divided into two-stage detectors (such as Faster R-CNN) and one-stage detectors (such as YOLO, SSD, RetinaNet), as well as the Transformer-based DETR series.
35 more timeline events
All Facts (20)
The YOLO (You Only Look Once) object detection algorithm was first proposed by Joseph Redmon in 2015.
75%VerifiedSince YOLO was proposed by Joseph Redmon in 2015, its core innovation has been to reformulate object detection as a regression problem solved by a single forward pass.
70%VerifiedYOLO treats object detection as a single regression problem, simultaneously predicting bounding boxes and class probabilities in a single forward pass.
70%VerifiedThe YOLO series of object detection models has evolved from V1 to V11.
65%VerifiedThe YOLO series of object detection models can already run in real time on embedded devices.
65%Unverified基于深度学习的目标检测方案（如YOLO）虽然泛化能力更强，但需要标注数据集和训练过程，对于UI元素固定的游戏场景来说过于重型
85%Unverified该项目使用OCR和图像识别替代传统的内存读取方式进行游戏状态感知，优势在于不需要逆向分析游戏内存结构，也不容易触发反作弊系统的检测
80%UnverifiedYOLOv8 has been officially renamed Ultralytics and shares the same author as YOLOv5.
60%UnverifiedYOLO models degrade sharply in performance when deployed across datasets, mainly because the annotation conventions of the two datasets differ (fixed-size bounding boxes vs. bounding boxes that fit the lesion contour).
50%UnverifiedThe updated content of this PhD's 13th large AI class includes the latest version of YOLO, the overall large model framework, and more.
50%UnverifiedRunning inference first with a pretrained model (such as the YOLO series or Faster R-CNN) can cover 60%-80% of common objects.
50%UnverifiedHuman detection is usually based on lightweight object detection models (such as the YOLO series or vendors' in-house models), running on the camera or on a local NVR.
50%UnverifiedRICO and VINS are both screenshot-domain data (pixel-level clear, no perspective distortion), and there is an obvious domain gap with the stroke noise and unclosed-line characteristics of hand-drawn wireframes, so when transferring directly it is recommended to use only their pretrained backbone networks and then fine-tune with a small number of hand-drawn samples.
50%UnverifiedAt a sample size of 200-500, YOLO-series models (such as YOLOv8n and YOLOv8s) can usually converge to usable accuracy.
50%UnverifiedAnnotating 200-500 hand-drawn wireframes with LabelImg or CVAT, combined with data augmentation (rotation, scaling, stroke perturbation), is enough to obtain a demo-level YOLO model that can be shown.
50%UnverifiedLightweight deep learning models such as YOLO can perform real-time object detection on edge devices.
50%UnverifiedFaster R-CNN is a two-stage detector, with high accuracy but slow speed; YOLO unifies detection into a regression problem, making it extremely fast; the difference between the two is essentially the trade-off between "find first, then judge" and "one step to the end," with the YOLO series being more popular in real-time scenarios, while two-stage methods still hold a place on accuracy-sensitive academic benchmarks.
50%UnverifiedMainstream 2D object detection methods can be divided into two-stage detectors (such as Faster R-CNN) and one-stage detectors (such as YOLO, SSD, RetinaNet), as well as the Transformer-based DETR series.
50%UnverifiedCNNs and their subsequent variants (such as ResNet, YOLO, and U-Net) form the skeleton of modern computer vision.
50%UnverifiedThe developer needs to count livestock on-device using a small YOLO/CoreML model, with no network, no account, and no server-side fallback.
50%