# YOLO-009: YOLOv4: Optimal Speed and Accuracy of Object Detection > Generated from a local paper card. Do not edit this page directly; edit the source card and rerun `vp run docs:generate`. > Internal wiki boundary: paper-card evidence and user-provided PDF downloads are available for private reading. Extracted full-paper text is not published. > Source: `library/paper_cards/YOLO-009.md` ## Paper Access * Internal PDF: Download Paper * DOI / official page: [needs confirmation](https://arxiv.org/abs/2004.10934) * Open-access page: [Open access page](https://arxiv.org/abs/2004.10934) * Deployment boundary: these PDF links are intended only for a private/protected internal 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: see Paper Access section above * 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?