Skip to content

Latest commit

 

History

History
61 lines (54 loc) · 3.14 KB

README_EN.md

File metadata and controls

61 lines (54 loc) · 3.14 KB
Raw

sophon-stream retinaface element

English | 简体中文

The sophon-stream retinaface element is a plugin in the sophon-stream framework that utilizes a combination of additional supervised (extra-supervised) and self-supervised learning (self-supervised) through multi-task learning. It is designed for pixel-level localization of faces of different sizes. This project provides a sample routine for this plugin; for details, please refer to retinaface Demo.

1. Features

  • Support for multiple video streams
  • Support for multi-threaded processing

2. Configuration Parameters

The sophon-stream retinaface plugin has some configurable parameters that can be set according to your needs. Here are some commonly used parameters:

{
    "configure": {
        "model_path": "../data/models/BM1684X/retinaface_mobilenet0.25_fp32_1b.bmodel",
        "max_face_count":50,
        "score_threshold":0.1,
        "threshold_nms": 0.4,
        "bgr2rgb": false,
        "mean": [
            104,
            117,
            123
        ],
        "std": [
            1,
            1,
            1
        ],
        "stage": [
            "pre"
        ]
    },
    "shared_object": "../../../build/lib/libretinaface.so",
    "name": "retinaface",
    "side": "sophgo",
    "thread_number": 1
}
Parameter Name Type Default Value Description
model_path String "../data/models/retinaface_mobilenet0.25_int8_1b.bmodel" Path to the retinaface model
max_face_count Integer 50 Maximum number of faces to detect
score_threshold Float 0.1 Confidence threshold for target detection
threshold_nms Float 0.4 NMS IOU threshold for target detection
bgr2rgb Bool false Whether to convert the image from BGR to RGB format; the default is BGR
mean Float Array None Mean values for image preprocessing, with a length of 3. The calculation is y=(x-mean)/std. If bgr2rgb=true, the order of the array should be R, G, B; otherwise, it should be B, G, R
std Float Array None Standard deviations for image preprocessing, with a length of 3. The calculation is the same as above. If bgr2rgb=true, the order of the array should be R, G, B; otherwise, it should be B, G, R
stage List ["pre"] Flags for the three stages of pre-processing, inference, and post-processing
shared_object String "../../../build/lib/libretinaface.so" Path to the libretinaface dynamic library
id Integer 0 Element ID
device_id Integer 0 TPU device number
name String "retinaface" Element name
side String "sophgo" Device type
thread_number Integer 1 Number of threads to start

Note:

  1. For the stage parameter, it needs to be set as one of "pre," "infer," "post," or a combination of adjacent items. These stages should be connected in the order of pre-processing, inference, and post-processing to elements. The purpose of allocating these three stages to three elements is to maximize the utilization of resources, enhancing the efficiency of detection.