NEWAgents can now see video via MCP.Try it now →
    Models/Detection & Recognition/deepinsight/retinaface-r50
    HFFace DetectionMIT

    retinaface-r50

    by deepinsight

    Single-stage face detection with landmark localization

    dl/month
    27Mparams
    Identifiers
    Model ID
    deepinsight/retinaface-r50
    Feature URI
    mixpeek://face_identity@v1/deepinsight_retinaface_r50

    Overview

    RetinaFace is a robust single-stage face detector that simultaneously performs face detection, 2D facial landmark localization, and 3D face reconstruction. It achieves state-of-the-art performance on WIDER FACE benchmark.

    On Mixpeek, RetinaFace powers face detection and identity matching across video and image content, enabling face-based search and filtering in retrieval pipelines.

    Architecture

    ResNet-50 backbone with multi-task learning heads: face classification, bounding box regression, 5-point landmark localization, and optional 3D face mesh regression. Uses Feature Pyramid Network (FPN) for multi-scale detection.

    Mixpeek SDK Integration

    import { Mixpeek } from "mixpeek";
    const mx = new Mixpeek({ apiKey: "API_KEY" });
    await mx.collections.ingest({
    collection_id: "my-collection",
    source: { url: "https://example.com/video.mp4" },
    feature_extractors: [{
    name: "face_detection",
    version: "v1",
    params: {
    model_id: "deepinsight/retinaface-r50"
    }
    }]
    });

    Capabilities

    • High-accuracy face detection across scales
    • 5-point facial landmark localization
    • Face embedding extraction for identity matching
    • Handles partial occlusion and extreme poses

    Use Cases on Mixpeek

    Media production, find all scenes featuring a specific person
    Security footage analysis, face-based identity search
    Photo library organization, auto-group by person

    Benchmarks

    DatasetMetricScoreSource
    WIDER Face (Hard)AP91.4%Deng et al., 2020 — Table 5
    WIDER Face (Easy)AP96.5%Deng et al., 2020 — Table 5

    Performance

    Input Size640×640 px
    GPU Latency~5ms / image (A100)
    CPU Latency~42ms / image
    GPU Throughput~200 images/sec (A100)
    GPU Memory~0.5 GB

    Specification

    FrameworkHF
    Organizationdeepinsight
    FeatureFace Detection
    Outputface embedding
    Modalitiesvideo, image
    RetrieverFace Filter
    Parameters27M
    LicenseMIT
    Downloads/mo

    Research Paper

    RetinaFace: Single-shot Multi-level Face Localisation in the Wild

    arxiv.org

    Build a pipeline with retinaface-r50

    Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

    Open Pipeline Builder