YOLOv3：一项渐进式改进

YOLOv3: An Incremental Improvement

约瑟夫·雷德蒙 Joseph Redmon · U. Washington · 2018-04-08 · arXiv:1804.02767 ↗ · 被引 25506

打开互动全文版（逐段中英对照 + 图/公式 + 论文问答）→

摘要 · Abstract

我们介绍了 YOLO 的一些更新！我们进行了一系列小的设计改进以提升性能。我们还训练了一个新的网络，效果相当不错。它比之前的版本稍大，但更准确。不过它仍然很快，别担心。在 320x320 分辨率下，YOLOv3 运行时间为 22 毫秒，mAP 为 28.2，与 SSD 精度相当但速度快三倍。当使用旧的 .5 IOU mAP 检测指标时，YOLOv3 表现相当出色。它在 Titan X 上以 51 毫秒达到 57.9 mAP@50，而 RetinaNet 需要 198 毫秒才能达到 57.5 mAP@50，性能相似但速度快 3.8 倍。一如既往，所有代码均可在 https://pjreddie.com/yolo/ 获取。

We present some updates to YOLO! We made a bunch of little design changes to make it better. We also trained this new network that's pretty swell. It's a little bigger than last time but more accurate. It's still fast though, don't worry. At 320x320 YOLOv3 runs in 22 ms at 28.2 mAP, as accurate as SSD but three times faster. When we look at the old .5 IOU mAP detection metric YOLOv3 is quite good. It achieves 57.9 mAP@50 in 51 ms on a Titan X, compared to 57.5 mAP@50 in 198 ms by RetinaNet, similar performance but 3.8x faster. As always, all the code is online at https://pjreddie.com/yolo/

核心贡献 · Key contributions

• 提出 YOLOv3，对 YOLO 进行增量改进，提升精度和速度。
  Proposes YOLOv3, an incremental improvement to YOLO with better accuracy and speed.
• 引入 Darknet-53，一种平衡效率与性能的新特征提取器。
  Introduces Darknet-53, a new feature extractor balancing efficiency and performance.
• 采用特征金字塔网络的多尺度预测，改进小目标检测。
  Adopts multi-scale predictions using feature pyramid networks for improved small object detection.
• 使用独立逻辑分类器进行多标签分类，避免 softmax 限制。
  Uses independent logistic classifiers for multilabel classification, avoiding softmax limitations.
• 在 COCO 上实现速度与精度的最优权衡，例如 51 毫秒内达到 57.9 mAP@50。
  Achieves state-of-the-art speed-accuracy trade-off on COCO, e.g., 57.9 mAP@50 in 51 ms.
• 开源代码和模型以促进可复现性。
  Open-sources code and model for reproducibility.

局限 · Limitations

• 在高 IOU 阈值（如 0.75）下性能显著下降，表明边界框对齐不佳。
  Performance drops significantly at higher IOU thresholds (e.g., 0.75), indicating poor box alignment.
• 在 COCO 中，中大型目标的性能不如小型目标。
  Underperforms on medium and large objects compared to small objects in COCO.
• 在 COCO 平均 AP（0.5:0.95）上未达到最佳结果，落后于 RetinaNet。
  Does not achieve top results on COCO average AP (0.5:0.95), lagging behind RetinaNet.
• 部分尝试的改进（如焦点损失、双 IOU 阈值）未带来收益。
  Some attempted improvements (e.g., focal loss, dual IOU thresholds) did not yield gains.

论文章节 · Sections（共 11）

• 摘要 Abstract
• 1 引言 1 Introduction
• 2 方法 2 The Deal
• 2.1 边界框预测 2.1 Bounding Box Prediction
• 2.2 类别预测 2.2 Class Prediction
• 2.3 多尺度预测 2.3 Predictions Across Scales
• 2.4 特征提取器 2.4 Feature Extractor
• 2.5 训练 2.5 Training
• 3 实验结果 3 How We Do
• 4 未成功的尝试 4 Things We Tried That Didn’t Work
• 5 结论 5 What This All Means

阅读逐段中英对照全文 →