# YOLO-004: Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks > 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-004.md` ## Paper Access * Internal PDF: Download Paper * DOI / official page: [10.1109/TPAMI.2016.2577031](https://doi.org/10.1109/TPAMI.2016.2577031) * Deployment boundary: these PDF links are intended only for a private/protected internal wiki. ## YOLO-004: Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks ## Metadata * ID: YOLO-004 * Title: Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks * Year: 2017 * DOI / URL: 10.1109/TPAMI.2016.2577031 * Local PDF: see Paper Access section above * Text artifact: local-only path withheld from docs site * Review status: `extracted` ## Study Type * Track: Object Detection / Region Proposals * Task: real-time object detection with learned region proposal networks * Participants or dataset: PASCAL VOC 2007/2012, MS COCO, ILSVRC/COCO competition settings; no human participants * Hardware: GPU-based CNN detection experiments * Channels or sensors: RGB images; no depth, EEG, or robot sensors ## Methods * Paradigm: two-stage detector that combines class-agnostic region proposals with Fast R-CNN detection * Signal processing or model: Region Proposal Network shares convolutional features with Fast R-CNN; anchor boxes cover multiple scales and aspect ratios * Training/calibration: alternating or approximate joint training is used to share features between RPN and detector * Online/offline: benchmarked as an object detection system; not a BCI or manipulation experiment ## Results * Metrics: detection mAP, proposal count, runtime/fps * Main findings: RPN makes proposals nearly cost-free relative to the detector; Faster R-CNN reaches 5 fps with VGG-16 and can run 5-17 fps depending on network; two-stage detection outperforms a one-stage dense-window variant in the reported comparison * Reported limitations: detects boxes and categories only; no grasp pose, hand-eye calibration, safety arbitration, or BCI integration ## Relevance To This Project * Supports: vision baseline for object candidate generation and a two-stage detector comparator against one-stage YOLO-style detectors * Conflicts with: Faster R-CNN is not a grasp planner and is slower than lightweight one-stage detectors used for real-time command-space updates * Design implication: use this card to position object detection choices and comparator families, not to justify direct robot grasp execution ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | Faster R-CNN combines an RPN with Fast R-CNN in a unified detector. | verified | The abstract and method sections state that RPN proposes object bounds/objectness and shares features with Fast R-CNN. | Abstract; Section 3 | | RPN proposals are designed to be low-cost because convolutional features are shared with the detector. | verified | The introduction and conclusion describe proposal computation as nearly cost-free once detector features are computed. | Sections 1, 5 | | Multi-scale and multi-aspect anchors are a central design feature. | verified | The RPN section defines anchors and reports default three scales and three aspect ratios. | Section 3.1.1 | | The reported detector reaches practical near-real-time rates on benchmarks. | verified | The abstract reports 5 fps with VGG-16 on GPU; the conclusion reports a 5-17 fps unified detection system. | Abstract; Section 5 | | Object detection outputs can seed a dynamic command space in SAH-BRI-Grasp. | inferred | Faster R-CNN verifies image-to-box/category detection, but the mapping from detected boxes to BCI-selectable commands is a project architecture choice. | Sections 3-4 | | Faster R-CNN does not verify grasp pose estimation or physical manipulation. | needs confirmation | The paper is an object detection benchmark, not a robotic grasping or calibration study. | Full paper scope | ## Open Questions * Which detector family should be used in experiments: YOLO for speed, Faster R-CNN as a two-stage comparator, or both? * How should SAH-BRI-Grasp handle detection flicker, target identity tracking, and box freezing before SSVEP presentation? * What confidence threshold and non-maximum suppression behavior should be logged as part of the command-space generator?