# YOLO-005: Microsoft COCO: Common Objects in Context > 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-005.md` ## Paper Access * Internal PDF: Download Paper * DOI / official page: [10.1007/978-3-319-10602-1\_48](https://doi.org/10.1007/978-3-319-10602-1_48) * Open-access page: [Open access page](https://arxiv.org/abs/1405.0312) * Deployment boundary: these PDF links are intended only for a private/protected internal wiki. ## YOLO-005: Microsoft COCO: Common Objects in Context ## Metadata * ID: YOLO-005 * Title: Microsoft COCO: Common Objects in Context * Year: 2014 * DOI / URL: 10.1007/978-3-319-10602-1\_48 * 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: large-scale object detection and instance segmentation dataset for natural-context scene understanding * Participants or dataset: MS COCO with 91 common object categories, 328,000 images, and 2.5 million labeled object instances; Amazon Mechanical Turk workers performed annotation tasks * Hardware: dataset and benchmark study; no robot, EEG, or closed-loop hardware is reported * Channels or sensors: RGB images from Flickr-style natural scenes; no depth sensor, robot sensor, or hand-eye calibration data is reported ## Methods * Paradigm: collection of mostly non-iconic everyday scenes with common objects in natural context, followed by category labeling, instance spotting, and per-instance segmentation * Signal processing or model: hierarchical AMT category labeling, instance spotting, modified segmentation UI, tight bounding boxes derived from masks, and DPMv5 bounding-box/segmentation baselines * Training/calibration: COCO split into 164,000 training images, 82,000 validation images, and 82,000 test images; DPMv5-C used COCO training examples; AMT segmentation workers completed category-specific training and periodic quality checks * Online/offline: offline dataset construction and benchmark analysis; no online BCI or robotic manipulation experiment ## Results * Metrics: dataset scale, categories and instances per image, DPM average precision, cross-dataset performance drop, and segmentation overlap * Main findings: COCO contains more object instances per image than ImageNet Detection and PASCAL VOC, emphasizes non-iconic/context-rich object views, and is more challenging than PASCAL VOC for DPM baselines * Reported limitations: the 2014 paper labels only "thing" categories, not "stuff"; future annotations suggested include occlusion, keypoints, scene types, attributes, and captions; the paper is not a grasping, BCI, or robotics study ## Relevance To This Project * Supports: using natural-scene object datasets to train or benchmark detectors that output object categories, boxes, and instance masks for scene-aware candidate generation * Conflicts with: COCO does not provide grasp poses, robot-frame coordinates, EEG/BCI interaction, target selection protocols, or hand-eye calibration evidence * Design implication: use COCO-backed detectors as visual candidate generators only; SAH-BRI-Grasp still needs target stability, BCI stimulus binding, grasp-pose estimation, calibration, and closed-loop validation ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | MS COCO is designed for object detection and segmentation in natural context rather than isolated object classification. | verified | The abstract and introduction frame COCO around complex everyday scenes, common objects in natural context, and instance-level segmentations. | Abstract; Introduction | | The 2014 COCO release described in this text contains 91 object categories, 328,000 images, and 2.5 million labeled instances. | verified | The abstract and introduction report 91 object types, 2.5 million labeled instances, and 328k images. | Abstract; Introduction | | COCO images contain richer object context than ImageNet Detection and PASCAL VOC in the reported comparison. | verified | The dataset statistics section reports an average of 3.5 categories and 7.7 instances per COCO image versus less than 2 categories and 3 instances per image for ImageNet/PASCAL. | Dataset Statistics | | The annotation pipeline uses category labeling, instance spotting, and instance segmentation with AMT workers. | verified | The image annotation section describes the three worker tasks and the quality-control process for segmentation workers. | Image Annotation | | COCO is more difficult than PASCAL VOC for the reported DPM detection baselines. | verified | The algorithmic analysis reports that DPM performance on COCO drops nearly by a factor of 2 relative to PASCAL and that overall performance is lower on COCO. | Algorithmic Analysis | | COCO-trained object detectors can support SAH-BRI-Grasp's scene-aware command candidate generator. | inferred | COCO supports natural-context object category/localization learning, but mapping detections into BCI-selectable commands is a project design choice. | Introduction; Dataset Statistics | | COCO does not verify robotic grasp execution, BCI target selection, or hand-eye calibration. | needs confirmation | The local text covers dataset construction and vision benchmarks, not robot manipulation or EEG interaction. | Full paper scope | ## Open Questions * Which COCO categories map cleanly to the tabletop objects expected in SAH-BRI-Grasp experiments? * How stable are COCO-trained detections under the project's camera viewpoint, lighting, occlusion, and object clutter? * What downstream module will convert selected object candidates into grasp poses and robot-frame targets? * Should the project use bounding boxes only, instance masks, or both when attaching SSVEP stimuli to detected objects?