DOCSEARCHbyAlgolia
Submit Docs

Hailo Model Zoo

repository·master·Indexed 20 days ago

https://github.com/hailo-ai/hailo_model_zoo

A collection of pre-trained deep learning models optimized for Hailo hardware. It provides tools for model parsing, optimization, compilation into HEF format, and accuracy evaluation. Features include the hailomz CLI for compilation and cascade evaluation, tools for converting TFRecord datasets to binary input files, and scripts for generating TFRecords from datasets such as ImageNet, COCO2017, Cityscapes, and WIDERFACE.

Tokens
127.9K
Snippets
209
Records
465
Agent score
71%

What's inside hailo_model_zoo

  1. Overview of the License Plate Detection model

    master

    The tiny_yolov4_license_plates model is a Tiny-YOLOv4 based network designed for detecting license plates on a single vehicle. It is trained to be robust across various weather conditions, lighting, vehicle types, and camera angles.

    Model Specifications:

    • Architecture: Tiny-YOLOv4
    • Parameters: 5.87M
    • GMACS: 3.4
    • Accuracy: 73.45 mAP (evaluated on an internal dataset of 5000 images)

    Input Requirements:

    • Format: RGB image
    • Size: 416x416x3
    • Normalization: Occurs on-chip

    Output Format:

    • Tensors: Two output tensors with sizes 13x13x18 and 26x26x18.
    • Structure: Each output contains 3 anchors. The 18 channels per anchor are a concatenation of 6 values:
      1. Bounding box center X
      2. Bounding box center Y
      3. Bounding box height
      4. Bounding box width
      5. Box objectness confidence score
      6. Class probability confidence score

  3. Overview of Person-ReID models

    master

    Hailo's Person Re-Identification (ReID) network is based on the RepVGG_A0 architecture. It is designed to work under various lighting conditions and camera angles. Two versions are available, distinguished by their embedding dimensions: 512 and 2048.

    Model Specifications

    Feature2048-dim Model512-dim Model
    ArchitectureRepVGG_A0RepVGG_A0
    Parameters9.65M7.68M
    GMACS0.890.89
    Rank1 (Market1501)89.8%89.3%

    Input/Output Requirements

    • Inputs: RGB images. Note that image resizing to 256x128x3 and image normalization both occur on-chip.
    • Outputs: A single embedding vector per query (either 2048 or 512 dimensions).

  4. Overview of the yolov5m_vehicles model

    master

    The yolov5m_vehicles model is a vehicle detection network based on the YOLOv5m architecture. It is trained for a single class (vehicles) and is designed to be robust across various weather conditions, lighting, and camera angles.

    Model Specifications

    • Architecture: YOLOv5m
    • Parameters: 21.47M
    • GMACS: 25.63
    • Accuracy: 46.0 mAP (evaluated on an internal dataset of 5000 images)

    Input Requirements

    • Format: RGB image
    • Size: 1080x1920x3
    • On-chip Processing: The Hailo hardware performs image resizing to 640x640x3 and image normalization on-chip.

    Output Format

    The model produces three output tensors with sizes of 20x20x18, 40x40x18, and 80x80x18. Each output contains 3 anchors. For each anchor, 6 values are concatenated into the 18 output channels:

    1. Bounding box coordinates (x center, y center, height, width)
    2. Box objectness confidence score
    3. Class probability confidence score

  5. Overview of License Plate Recognition (LPRNet)

    master
    Hailo's License Plate Recognition network (lprnet) is a convolutional network designed to predict registration numbers from license plates under various weather and lighting conditions. It is based on the LPRNet architecture with several modifications, including a ResNet-like backbone with 4 stages (each containing 2 residual blocks) and support for a maximal license plate length of 19 digits.

  6. Overview of Public Models and Formats

    master

    Hailo provides pre-trained models in multiple formats to support different stages of the AI workflow and deployment on Hailo hardware:

    • ONNX / TF formats: Used for pre-trained models typically used during development or training.
    • HEF (Hailo Executable Format): Pre-compiled binary files used for direct execution on Hailo devices.

    Models are categorized by Task Type (e.g., Object Detection, Classification, Segmentation) and are optimized for specific Hailo hardware families, including Hailo-10H, Hailo-15H, Hailo-15L, Hailo-8, and Hailo-8L.

  7. Understand Semantic Segmentation model availability for HAILO15L

    master

    The HAILO15L platform supports several semantic segmentation models trained on different datasets including Pascal VOC, Cityscapes, and Oxford-IIIT Pet.

    Models are provided in several formats and resources:

    • HEF: Compiled models in RGB format.
    • NV12: Compiled models in NV12 format.
    • RGBX: Compiled models in RGBX format.
    • PT (Pretrained): Downloadable pretrained model files in ZIP format.
    • S (Source): Links to the original open-source model repositories.
    • PR (Profiler Report): Performance profiling reports for the compiled models.
    • ⭐: Indicates networks used by hailo-apps infrastructure.

  8. Understand Super Resolution model availability and formats for HAILO15L

    master

    The Hailo Model Zoo provides several Super Resolution models optimized for the HAILO15L platform. These models are categorized by their dataset benchmarks (Bsd100 and Div2K) and are available in various formats for different deployment needs.

    Available Model Formats:

    • HEF: Compiled models in RGB format.
    • NV12: Compiled models in NV12 format.
    • RGBX: Compiled models in RGBX format.
    • PT (Pretrained): Downloadable pretrained model files in ZIP format.
    • S (Source): Links to the original open-source model repositories.
    • PR (Profiler Report): Performance profiling reports for the compiled models.

    Measurement Context: Benchmarks are measured from the SoC at room temperature using Hailo Dataflow Compiler version v5.4.0.

  9. Understand Instance Segmentation model availability for HAILO-15L

    master

    The Hailo Model Zoo provides several pretrained instance segmentation models optimized for the HAILO-15L hardware. These models are trained on the COCO dataset and are available in various formats for deployment.

    Key model families include:

    • YOLACT: yolact_regnetx_1.6gf, yolact_regnetx_800mf
    • YOLOv5 Segmentation: yolov5l_seg, yolov5m_seg, yolov5s_seg, yolov5n_seg
    • YOLOv8 Segmentation: yolov8n_seg, yolov8s_seg

    Models marked with ⭐ are specifically supported by hailo-apps infrastructure.

  10. Available architectures for retraining

    master

    The Hailo Model Zoo supports retraining for various computer vision tasks. Each task has specific documentation located within the training/ directory of the repository.

    Object Detection

    • YOLOv5
    • YOLOv8
    • YOLOX
    • DAMO-YOLO
    • NanoDet

    Pose Estimation

    • CenterPose

    Single Person Pose Estimation

    • MSPN

    Semantic Segmentation

    • FCN

    Instance Segmentation

    • YOLACT
    • YOLOv8_seg

    Face Recognition

    • ArcFace

  11. Understand Pose Estimation model availability for HAILO15H

    master

    The Hailo Model Zoo provides several pose estimation models optimized for the HAILO15H platform. These models are categorized by their performance metrics (AP, FPS) and input resolutions.

    Key model types available:

    • centerpose_regnetx_800mf: High resolution (512x512x3), moderate FPS.
    • yolov8m_pose: Medium scale, optimized for Hailo-apps usage.
    • yolov8s_pose: Small scale, high FPS, optimized for Hailo-apps usage.

    Models are provided in various formats including HEF (RGB), NV12, and RGBX.

  12. Understand Zero-Shot Depth Estimation models for HAILO8L

    master

    The Hailo Model Zoo provides pretrained and compiled Zero-Shot Depth Estimation models optimized for the HAILO8L hardware. These models can be used for estimating depth from single RGB images without specific training for the target scene.

    Available Models

    Network NameInput ResolutionParams (M)OPS (G)FPS (BS=1)FPS (BS=8)
    depth_anything_v2_vits224x224x324.216.732.491.4
    depth_anything_vits224x224x324.216.737.498.6

    Measurement Context

    Performance metrics (FPS and AbsRel) were measured under the following conditions:

    • System host: Intel® Core™ i5-9400 CPU @ 2.90GHz
    • Hailo Dataflow Compiler Version: v2.19.0
    • Measurement conditions: PCIe Gen 3 x 4 lanes, room temperature