# YOLO-001: You Only Look Once: Unified, Real-Time Object Detection > 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-001.md` ## 论文访问 * 内部 PDF: 下载论文 PDF * DOI / 官方页面: [10.1109/CVPR.2016.91](https://doi.org/10.1109/CVPR.2016.91) * 开放访问页面: [Open access page](https://arxiv.org/abs/1506.02640) * 部署边界: 这些 PDF 链接只适合私有/受保护的内部 wiki。 ## YOLO-001: You Only Look Once: Unified, Real-Time Object Detection ## Metadata * ID: YOLO-001 * Title: You Only Look Once: Unified, Real-Time Object Detection * Year: 2016 * DOI / URL: 10.1109/CVPR.2016.91 * 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 * Participants or dataset: PASCAL VOC 2007/2012 detection, ImageNet pretraining, and Picasso/People-Art cross-domain person detection * Hardware: Titan X GPU; webcam demo is discussed * Channels or sensors: RGB image input; no depth sensor, robot sensor, or hand-eye calibration ## Methods * Paradigm: object detection as a single regression problem from whole image to bounding boxes, class probabilities, and confidence scores * Signal processing or model: YOLO one-stage CNN with 24 convolutional layers and 2 fully connected layers; Fast YOLO with 9 convolutional layers; VOC output grid S=7, B=2, C=20 * Training/calibration: ImageNet pretraining, detection training for about 135 epochs, non-maximum suppression after prediction * Online/offline: real-time detection evaluation and webcam demo; no robotic closed-loop experiment ## Results * Metrics: YOLO 45 FPS and 63.4% mAP on VOC 2007; Fast YOLO 155 FPS and 52.7% mAP; YOLO 57.9% mAP on VOC 2012 * Main findings: a unified one-stage detector can provide real-time class-labeled bounding boxes and confidence scores * Reported limitations: localization error is a major failure mode; small objects, groups of objects, and unusual aspect ratios/configurations are weaker cases; accuracy trails some slower state-of-the-art detectors ## Relevance To This Project * Supports: using detected object class, bounding box, and confidence as inputs to a dynamic command-space candidate manager * Conflicts with: YOLO alone does not produce grasp pose, robot coordinates, target affordances, or hand-eye calibration * Design implication: use YOLO for object-candidate generation only; downstream modules must handle tracking/freeze, grasp synthesis, calibration, confirmation, and safety gating ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | YOLO can produce real-time object class, bounding-box, and confidence outputs. | verified | The paper formulates object detection as a single regression problem and reports real-time FPS/mAP results. | Abstract; Unified Detection; Experiments | | YOLO has localization and small-object limitations. | verified | The limitations section identifies localization error, small objects, groups, and unusual aspect ratios as weaknesses. | Limitations | | YOLO outputs can seed a scene-aware BCI command space. | inferred | The paper supplies visual candidate primitives, but does not discuss BCI or command-space interfaces. | Unified Detection | | Object detection is not equivalent to grasp-pose estimation. | verified | The paper distinguishes its multi-object bbox/class detection task from MultiGrasp-style graspable-region prediction. | MultiGrasp comparison | | Hand-eye calibration is outside YOLO-001's scope. | needs confirmation | No robot frame, camera-to-arm transform, or calibration method is provided. | Whole paper | ## Open Questions * How should detection confidence, class, and box stability be combined into BCI-selectable candidates? * How long should object candidates be frozen before SSVEP stimulation? * Which grasp-pose module should consume the selected object after YOLO candidate generation?