# YOLO-007: Focal Loss for Dense 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-007.md` ## Paper Access * Internal PDF: Download Paper * DOI / official page: [10.1109/ICCV.2017.324](https://doi.org/10.1109/ICCV.2017.324) * Open-access page: [Open access page](https://arxiv.org/abs/1708.02002) * Deployment boundary: these PDF links are intended only for a private/protected internal wiki. ## YOLO-007: Focal Loss for Dense Object Detection ## Metadata * ID: YOLO-007 * Title: Focal Loss for Dense Object Detection * Year: 2017 * DOI / URL: 10.1109/ICCV.2017.324 * 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: Dense one-stage object detection and foreground/background class-imbalance mitigation. * Participants or dataset: COCO trainval35k, minival, and test-dev. * Hardware: Multi-GPU detector training and detector timing comparisons. * Channels or sensors: RGB image object detection. ## Methods * Paradigm: One-stage dense detector benchmark. * Signal processing or model: RetinaNet with ResNet-FPN pyramid levels, dense anchors, separate classification and box-regression subnets, sigmoid classification, and focal loss `FL = -alpha_t(1 - p_t)^gamma log(p_t)`. * Training/calibration: ResNet-50/101-FPN backbones pretrained on ImageNet1k; SGD over 8 GPUs; final models use longer training and scale jitter. * Online/offline: Offline detector benchmark with speed/accuracy comparison. ## Results * Metrics: COCO AP, AP50, AP75, and inference time. * Main findings: Alpha-balanced cross-entropy reaches at most 31.1 AP, while focal loss reaches 34.0 AP in matched ablations. The focal-loss baseline reaches 36.0 AP versus best OHEM at 32.8 AP. RetinaNet-101-800 reaches 37.8 AP at 198 ms; the longer/scale-jitter model reaches 39.1 AP, 59.1 AP50, and 42.3 AP75. * Reported limitations: RetinaNet improves dense-detector accuracy but is slower than high-frame-rate YOLO-style detectors; it is not a BCI, grasping, or robot-control method. ## Relevance To This Project * Supports: Accuracy-focused one-stage detector comparator and class-imbalance handling. * Conflicts with: Its latency may be too high for live candidate refresh, and the paper does not study BCI command spaces or grasp-pose execution. * Design implication: Detector choice should be evaluated as a speed/stability/false-positive tradeoff, not as a pure AP ranking. ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | Focal loss improves dense one-stage detector accuracy under class imbalance. | verified | Matched ablation improves from 31.1 AP with alpha-balanced CE to 34.0 AP with focal loss. | Focal-loss ablation | | RetinaNet is an accuracy-focused one-stage comparator. | verified | RetinaNet-101-800 reaches 37.8 AP at 198 ms, and longer/scale-jitter training reaches 39.1 AP on COCO test-dev. | COCO result table | | RetinaNet can inform detector comparison but not robot control claims. | inferred | The paper contains no EEG, BCI, SSVEP stimulus binding, grasping, or shared-autonomy evaluation. | Project synthesis | ## Open Questions * Is RetinaNet latency acceptable for live `SAH-BRI-Grasp` candidate generation? * Does focal loss improve the actual tabletop object distribution used in this project? * How should AP, AP50, false positives, and temporal stability be weighted for BCI-selection safety?