# YOLO-010: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors > 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-010.md` ## 论文访问 * 内部 PDF: 下载论文 PDF * DOI / 官方页面: [needs confirmation](https://arxiv.org/abs/2207.02696) * 开放访问页面: [Open access page](https://arxiv.org/abs/2207.02696) * 部署边界: 这些 PDF 链接只适合私有/受保护的内部 wiki。 ## YOLO-010: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors ## Metadata * ID: YOLO-010 * Title: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors * Year: 2022 * DOI / URL: https://arxiv.org/abs/2207.02696 * 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 detector architecture, scaling, re-parameterization, and trainable bag-of-freebies evaluation. * Participants or dataset: MS COCO only; train2017 for training, val2017 for hyperparameter selection, and test2017 for final comparison. * Hardware: Real-time detector timing reported on V100. * Channels or sensors: RGB image object detection. ## Methods * Paradigm: Real-time one-stage detector benchmark. * Signal processing or model: YOLOv7 family with E-ELAN, compound scaling for concatenation-based models, planned re-parameterized convolution, auxiliary heads, lead-guided/coarse-to-fine label assignment, BN fusion, YOLOR implicit knowledge, and EMA inference model. * Training/calibration: Trained from scratch without external datasets or pretrained weights. * Online/offline: Offline detector benchmark with real-time inference reporting. ## Results * Metrics: COCO AP and FPS. * Main findings: The abstract reports 56.8% AP among real-time detectors at 30+ FPS on V100. Reported configurations include YOLOv7 at 51.4% AP / 161 FPS, YOLOv7-X at 53.1% AP / 114 FPS, YOLOv7-W6 at 54.9% AP / 84 FPS, YOLOv7-E6 at about 55.9%-56.0% AP / 56 FPS, and YOLOv7-E6E at 56.8% AP / 36 FPS. Compared with YOLOv4, YOLOv7 reports 75% fewer parameters, 36% less computation, and +1.5 AP. * Reported limitations: Strong detector benchmark evidence, but no BCI loop, grasp affordance output, robot grasping, or project-hardware validation. ## Relevance To This Project * Supports: A stronger modern YOLO-family detector comparator for low-latency object candidate generation. * Conflicts with: COCO benchmark performance does not prove scene-specific target stability, SSVEP decodability, or grasp success. * Design implication: YOLOv7 can be a reasonable detector candidate, but `SAH-BRI-Grasp` must validate target hardware speed, custom object classes, candidate stability, and downstream grasp handling. ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | YOLOv7 reports high AP at real-time speed on COCO. | verified | YOLOv7-E6E reports 56.8% AP at 36 FPS, while YOLOv7 reports 51.4% AP at 161 FPS. | Abstract / result tables | | YOLOv7 improves the YOLO detector-family comparison point. | verified | The paper reports 75% fewer parameters, 36% less computation, and +1.5 AP compared with YOLOv4. | Detector comparison discussion | | YOLOv7 can support perception-front-end selection but not BCI/grasp claims. | inferred | The study is a detector benchmark with no EEG, SSVEP, robot grasping, or shared-autonomy evaluation. | Project synthesis | ## Open Questions * Is YOLOv7 or a newer local detector variant best for the target hardware? * Do project object classes require fine-tuning? * How will detections be frozen, tracked, and handed to grasp planning during EEG decision windows?