We introduce YOLO9000, a state-of-the-art, real-time object detection system that can detect over 9000 object categories. First we propose various improvements to the YOLO detection method, both novel and drawn from prior work. The improved model, YOLOv2, is state-of-the-art on standard detection tasks like PASCAL VOC and COCO. At 67 FPS, YOLOv2 gets 76.8 mAP on VOC 2007. At 40 FPS, YOLOv2 gets 78.6 mAP, outperforming state-of-the-art methods like Faster RCNN with ResNet and SSD while still running significantly faster. Finally we propose a method to jointly train on object detection and classification. Using this method we train YOLO9000 simultaneously on the COCO detection dataset and the ImageNet classification dataset. Our joint training allows YOLO9000 to predict detections for object classes that don't have labelled detection data. We validate our approach on the ImageNet detection task. YOLO9000 gets 19.7 mAP on the ImageNet detection validation set despite only having detection data for 44 of the 200 classes. On the 156 classes not in COCO, YOLO9000 gets 16.0 mAP. But YOLO can detect more than just 200 classes; it predicts detections for more than 9000 different object categories. And it still runs in real-time.
核心贡献 · Key contributions
提出 YOLOv2，融合批归一化、高分辨率分类器、锚框、维度聚类、直接位置预测、细粒度特征和多尺度训练。 Proposes YOLOv2 with batch normalization, high-resolution classifier, anchor boxes, dimension clusters, direct location prediction, fine-grained features, and multi-scale training.
引入 WordTree 层次化标签结构，联合 ImageNet 和 COCO 进行分类与检测训练。 Introduces WordTree, a hierarchical label structure combining ImageNet and COCO for joint classification and detection training.
开发 YOLO9000，通过检测与分类数据联合训练，实时识别超过 9000 个类别。 Develops YOLO9000, a real-time detector recognizing over 9000 categories via joint training on detection and classification data.
YOLOv2 在 PASCAL VOC 和 COCO 上达到最优，速度和精度均超越 Faster R-CNN 和 SSD。 Achieves state-of-the-art results on PASCAL VOC and COCO with YOLOv2, outperforming Faster R-CNN and SSD in speed and accuracy.
在 ImageNet 检测任务上仅用 44/200 类检测标签即达 19.7 mAP，证明对未见类的泛化能力。 Demonstrates 19.7 mAP on ImageNet detection with only 44 of 200 classes having detection labels, proving generalization to unseen classes.
多尺度训练使单个 YOLOv2 模型可在 288x288 至 608x608 分辨率间权衡速度与精度。 Provides multi-scale training enabling a single YOLOv2 model to trade off speed and accuracy across resolutions from 288x288 to 608x608.
局限 · Limitations
YOLOv2 对小物体性能有限，尽管有细粒度特征，但特征图仍较粗（13x13）。 YOLOv2's performance on small objects is limited due to coarse 13x13 feature map despite fine-grained features.
YOLO9000 对服装和设备等类别表现不佳，因 COCO 中缺乏这些类的检测数据。 YOLO9000 struggles with categories like clothing and equipment due to lack of detection data for those classes in COCO.
联合训练对分类图像使用弱标签，可能引入噪声并限制检测精度。 Joint training assumes weak labels for classification images, which may introduce noise and limit detection accuracy.
WordTree 层次结构基于 ImageNet 构建，可能无法泛化到其他标签结构不同的数据集。 WordTree hierarchy is built from ImageNet and may not generalize to other datasets with different label structures.
该方法依赖过采样 COCO 平衡数据集大小，可能无法扩展到更大的分类数据集。 The method relies on oversampling COCO to balance dataset sizes, which may not scale to even larger classification datasets.