# YOLO-003: YOLOv3: An Incremental Improvement > 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-003.md` ## 论文访问 * 内部 PDF: 下载论文 PDF * DOI / 官方页面: [needs confirmation](https://arxiv.org/abs/1804.02767) * 开放访问页面: [Open access page](https://arxiv.org/abs/1804.02767) * 部署边界: 这些 PDF 链接只适合私有/受保护的内部 wiki。 ## YOLO-003: YOLOv3: An Incremental Improvement ## Metadata * ID: YOLO-003 * Title: YOLOv3: An Incremental Improvement * Year: 2018 * DOI / URL: https://arxiv.org/abs/1804.02767 * 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: Incremental improvements to low-latency one-stage object detection. * Participants or dataset: COCO detection evaluation and ImageNet classifier evaluation for Darknet-53. * Hardware: Titan X-class GPU timing is reported. * Channels or sensors: RGB image object detection. ## Methods * Paradigm: One-stage detector benchmark with multi-scale prediction. * Signal processing or model: YOLOv3 uses anchor dimension clusters, logistic objectness, independent logistic multilabel class prediction instead of softmax, three-scale prediction with feature reuse, and Darknet-53 residual feature extraction. * Training/calibration: Full-image training, multi-scale training, data augmentation, batch normalization, and Darknet training stack. * Online/offline: Offline benchmark with low-latency inference timing. ## Results * Metrics: COCO AP, AP50, and inference time. * Main findings: YOLOv3-320 reports 28.2 COCO AP in 22 ms; YOLOv3-416 reports 31.0 AP in 29 ms; YOLOv3-608 reports 33.0 AP and 57.9 AP50 in 51 ms. The paper emphasizes AP50 strength relative to RetinaNet with much lower latency. * Reported limitations: YOLOv3 is weaker on COCO AP across IoU .5:.95 than on AP50, suggesting less precise box alignment. The paper notes failed experiments including focal loss, linear x/y prediction, and dual IoU threshold assignment. ## Relevance To This Project * Supports: A legacy real-time detector comparator for dynamic object-command generation where latency matters. * Conflicts with: AP50-style localization may be sufficient for visual selection but is not evidence for grasp-pose precision or safe robot execution. * Design implication: If YOLO-like detectors are used, `SAH-BRI-Grasp` should log both detector latency and candidate-box stability, not only task success. ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | YOLOv3 provides low-latency detector performance. | verified | YOLOv3-320 reports 28.2 AP in 22 ms; YOLOv3-416 reports 31.0 AP in 29 ms; YOLOv3-608 reports 33.0 AP in 51 ms. | Abstract / experiments | | YOLOv3 is stronger on AP50 than strict COCO AP. | verified | The paper reports 57.9 AP50 at 51 ms and discusses AP50 strength relative to RetinaNet. | COCO result discussion | | Detector speed supports candidate refresh, but not grasp execution. | inferred | The paper evaluates object detection only, with no BCI, grasp affordance, or robot-control loop. | Project synthesis | ## Open Questions * Is AP50-level box localization adequate for SSVEP target overlays? * What confidence, non-max suppression, and temporal smoothing settings should define selectable candidates? * Does target hardware match the reported timing regime? * How should candidate refresh be paused during EEG decision windows?