We present YOLO, a new approach to object detection. Prior work on object detection repurposes classifiers to perform detection. Instead, we frame object detection as a regression problem to spatially separated bounding boxes and associated class probabilities. A single neural network predicts bounding boxes and class probabilities directly from full images in one evaluation. Since the whole detection pipeline is a single network, it can be optimized end-to-end directly on detection performance. Our unified architecture is extremely fast. Our base YOLO model processes images in real-time at 45 frames per second. A smaller version of the network, Fast YOLO, processes an astounding 155 frames per second while still achieving double the mAP of other real-time detectors. Compared to state-of-the-art detection systems, YOLO makes more localization errors but is far less likely to predict false detections where nothing exists. Finally, YOLO learns very general representations of objects. It outperforms all other detection methods, including DPM and R-CNN, by a wide margin when generalizing from natural images to artwork on both the Picasso Dataset and the People-Art Dataset.
核心贡献 · Key contributions
将目标检测建模为从图像像素到边界框和类别概率的单一回归问题。 Frames object detection as a single regression problem from image pixels to bounding boxes and class probabilities.
统一架构支持端到端训练，YOLO 实现 45 fps、Fast YOLO 实现 155 fps 的实时速度。 Unified architecture enables end-to-end training and real-time speeds at 45 fps for YOLO and 155 fps for Fast YOLO.
YOLO 对图像进行全局推理，背景误检率比 Fast R-CNN 降低一半。 YOLO reasons globally about the image, reducing background false positives by half compared to Fast R-CNN.
学习可泛化的物体表征，从自然图像泛化到艺术作品时优于 DPM 和 R-CNN。 Learns generalizable object representations, outperforming DPM and R-CNN when generalizing from natural images to artwork.
将 YOLO 与 Fast R-CNN 结合，通过减少背景错误在 VOC 2007 上将 mAP 提升 3.2%。 Combining YOLO with Fast R-CNN boosts mAP by 3.2% on VOC 2007 by reducing background errors.
开源框架并提供预训练模型，便于复现和部署。 Open-source framework with pretrained models, enabling easy reproduction and deployment.
局限 · Limitations
YOLO 比最先进的检测器产生更多定位错误，尤其是小物体。 YOLO makes more localization errors than state-of-the-art detectors, especially for small objects.
空间约束限制了 YOLO 可预测的邻近物体数量，难以处理小物体群。 Spatial constraints limit the number of nearby objects YOLO can predict, struggling with groups of small objects.
损失函数对大框和小框的错误一视同仁，与检测性能不完全对齐。 Loss function treats errors equally for large and small boxes, not aligning perfectly with detection performance.
下采样层的粗粒度特征阻碍了精确的边界框预测。 Coarse features from downsampling layers hinder precise bounding box prediction.
YOLO 在 VOC 2012 等标准基准上的准确率落后于最先进方法。 YOLO lags behind state-of-the-art in accuracy on standard benchmarks like VOC 2012.
论文章节 · Sections（共 16）
摘要Abstract
1 引言1 Introduction
2 统一检测2 Unified Detection
2.1 网络设计2.1 Network Design
2.2 训练2.2 Training
2.3 推理2.3 Inference
2.4 YOLO 的局限性2.4 Limitations of YOLO
3 与其他检测系统的比较3 Comparison to Other Detection Systems
4 实验4 Experiments
4.1 与其他实时系统的比较4.1 Comparison to Other Real-Time Systems
4.2 VOC 2007 错误分析4.2 VOC 2007 Error Analysis
4.3 结合 Fast R-CNN 和 YOLO4.3 Combining Fast R-CNN and YOLO
4.4 VOC 2012 结果4.4 VOC 2012 Results
4.5 泛化能力：艺术品中的人物检测4.5 Generalizability: Person Detection in Artwork