# YOLO-002: YOLO9000: Better, Faster, Stronger > 本页由本地论文卡片生成。请不要直接编辑本页；修改源卡片后运行 `vp run docs:generate`。 > 中文支持说明：本页是中文站点镜像，保留论文标题、证据状态和来源字段的原始表述，避免未经核实的翻译改变证据边界。 > 内部 wiki 边界：本页提供论文卡片证据和用户提供的 PDF 下载入口，仅用于私有阅读；全文抽取文本不发布。 > 来源: `library/paper_cards/YOLO-002.md` ## 论文访问 * 内部 PDF: 下载论文 PDF * DOI / 官方页面: [10.1109/CVPR.2017.690](https://doi.org/10.1109/CVPR.2017.690) * 开放访问页面: [Open access page](https://arxiv.org/abs/1612.08242) * 部署边界: 这些 PDF 链接只适合私有/受保护的内部 wiki。 ## YOLO-002: YOLO9000: Better, Faster, Stronger ## Metadata * ID: YOLO-002 * Title: YOLO9000: Better, Faster, Stronger * Year: 2017 * DOI / URL: 10.1109/CVPR.2017.690 * Local PDF: 见上方论文访问区块 * Text artifact: local-only path withheld from docs site * Review status: `extracted` ## Study Type * Track: YOLO / Detection * Task: Real-time object detection and joint detection/classification scaling. * Participants or dataset: PASCAL VOC 2007/2012, COCO test-dev2015, ImageNet detection validation, and joint COCO/ImageNet training. * Hardware: Detector benchmark compute; not a BCI or robot-hardware study. * Channels or sensors: RGB image object detection. ## Methods * Paradigm: One-stage object detection with speed/accuracy comparison and large-vocabulary detector training. * Signal processing or model: YOLOv2 with batch normalization, high-resolution classifier pretraining, anchor boxes, k-means dimension priors, direct location prediction, passthrough features, multi-scale training, and Darknet-19; YOLO9000 with WordTree hierarchical classification and joint detection/classification training. * Training/calibration: Detection training on PASCAL/COCO and joint COCO detection plus ImageNet classification training; the combined WordTree has 9418 classes. * Online/offline: Offline detector benchmark with real-time speed reporting. ## Results * Metrics: mAP, AP50, FPS. * Main findings: YOLOv2 reports 76.8 mAP on VOC 2007 at 67 FPS and 78.6 mAP at 40 FPS; VOC2012 mAP is 73.4; COCO test-dev reports 21.6 AP and 44.0 AP50; YOLO9000 reports 19.7 mAP on ImageNet detection validation and 16.0 mAP on classes without COCO detection labels. * Reported limitations: Original YOLO had localization errors and relatively low recall; YOLO9000 struggled with categories lacking COCO box labels, such as some clothing/equipment categories. ## Relevance To This Project * Supports: Real-time class, bounding-box, and confidence output as a basis for scene candidate generation. * Conflicts with: Does not provide BCI target selection, SSVEP stimulus binding, grasp-pose prediction, or robot safety evidence. * Design implication: YOLO-family outputs can seed a dynamic command space, but `SAH-BRI-Grasp` must separately handle target freezing, false detections, confirmation, and grasp planning. ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | YOLOv2 supports real-time object detection with bounding boxes and classes. | verified | The paper reports 76.8 mAP at 67 FPS and 78.6 mAP at 40 FPS on VOC 2007. | Abstract / experiments | | Joint detection/classification can expand detector vocabulary. | verified | YOLO9000 uses WordTree to combine COCO detection data and ImageNet classification labels across 9418 classes. | WordTree / joint training sections | | YOLO evidence supports candidate generation, not executable grasp commands. | inferred | The paper is an object-detection benchmark and contains no BCI, SSVEP, grasp-pose, or robot execution study. | Project synthesis | ## Open Questions * Are YOLOv2/YOLO9000 boxes stable enough for flickering SSVEP overlays? * What candidate confidence and temporal-smoothing rules should freeze a command space? * How should false positives, missed objects, and class ambiguity affect confirmation or cancellation? * Which separate grasp-pose module should consume selected object candidates?