# YOLO-009: YOLOv4: Optimal Speed and Accuracy of 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-009.md`

## 论文访问

* 内部 PDF: <a href={"/papers/YOLO-009.pdf"} download style={{ display: "inline-flex", alignItems: "center", justifyContent: "center", minHeight: "2.25rem", padding: "0.45rem 0.8rem", borderRadius: "6px", backgroundColor: "#047857", color: "#ffffff", fontWeight: 700, lineHeight: 1, textDecoration: "none", boxShadow: "0 1px 2px rgba(15, 23, 42, 0.22)" }}>下载论文 PDF</a>
* DOI / 官方页面: [needs confirmation](https://arxiv.org/abs/2004.10934)
* 开放访问页面: [Open access page](https://arxiv.org/abs/2004.10934)
* 部署边界: 这些 PDF 链接只适合私有/受保护的内部 wiki。

## YOLO-009: YOLOv4: Optimal Speed and Accuracy of Object Detection

## Metadata

* ID: YOLO-009
* Title: YOLOv4: Optimal Speed and Accuracy of Object Detection
* Year: 2020
* DOI / URL: https://arxiv.org/abs/2004.10934
* 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 detector architecture and training-strategy evaluation.
* Participants or dataset: ImageNet/ILSVRC2012 classifier ablations and MS COCO test-dev 2017 detector experiments.
* Hardware: Training designed for conventional GPU use; speed reported on Tesla V100.
* Channels or sensors: RGB image object detection.

## Methods

* Paradigm: Real-time one-stage detector benchmark with extensive bag-of-freebies and bag-of-specials ablations.
* Signal processing or model: YOLOv4 uses CSPDarknet53 backbone, SPP, PAN, and YOLOv3 anchor-based head, plus techniques including Mosaic, self-adversarial training, CmBN, DropBlock, CIoU loss, Mish, SAM, PAN, and DIoU-NMS.
* Training/calibration: MS COCO detector training with batch-size and mini-batch comparisons; many ablations use one GPU.
* Online/offline: Offline detector benchmark with real-time speed reporting.

## Results

* Metrics: COCO AP, AP50, AP75, and FPS.
* Main findings: YOLOv4 reports 43.5% AP and 65.7% AP50 on MS COCO, with about 65 FPS on Tesla V100 in the abstract; comparison tables report YOLOv4-608 at 62 FPS, 43.5 AP, 65.7 AP50, and 47.3 AP75. Mini-batch 4 versus 8 with BoF/BoS gives nearly identical detector AP.
* Reported limitations: Evidence is detector benchmarking, not BCI interaction, robotic grasping, grasp affordance prediction, or camera-to-robot latency validation.

## Relevance To This Project

* Supports: A mature real-time YOLO-family detector option for object-candidate generation.
* Conflicts with: Does not validate detection-to-grasp transfer or SSVEP target stability.
* Design implication: If used as a perception front end, YOLOv4 needs project-specific latency, object-class, and candidate-stability validation.

## Extracted Evidence

| Claim | Status | Evidence Note | Page/Section |
| --- | --- | --- | --- |
| YOLOv4 improves real-time detector speed/accuracy tradeoffs on COCO. | verified | Reports 43.5% AP and 65.7% AP50, with about 65 FPS on Tesla V100. | Abstract / detector comparison |
| YOLOv4 combines backbone, neck, head, and training tricks into a real-time detector. | verified | The paper specifies CSPDarknet53, SPP, PAN, YOLOv3 head, and BoF/BoS methods including Mosaic, SAT, CmBN, DropBlock, CIoU, and DIoU-NMS. | Method sections |
| YOLOv4 suitability for SAH-BRI-Grasp remains an engineering inference. | inferred | The paper evaluates object detection only and does not include BCI, SSVEP, grasp affordance, or robot experiments. | Project synthesis |

## Open Questions

* What latency and AP are achieved on the project computer/camera stream?
* Which project objects require fine-tuning beyond COCO labels?
* How should YOLO boxes be mapped to stable SSVEP-selectable candidates and separate grasp regions?