retinaface-resnet50-pytorch

Use Case and High-Level Description

The retinaface-resnet50-pytorch model is a PyTorch* implementation of medium size RetinaFace model with ResNet50 backbone for Face Localization. It can output face bounding boxes and five facial landmarks in a single forward pass. More details provided in the paper and repository

Specification

Metric

Value

AP (WIDER)

91.78%

GFLOPs

88.8627

MParams

27.2646

Source framework

PyTorch*

Average Precision (AP) is defined as an area under the precision/recall curve. All numbers were evaluated by taking into account only faces bigger than 64 x 64 pixels.

Accuracy validation approach different from described in the original repository. In contrast to the Accuracy Checker strategy where whole set is evaluated, the validation set is divided into 3 predefined subsets(hard, medium and easy) and all subsets are verified separately in the original evaluation strategy. For details about original WIDER results please see repository.

Input

Original model:

Image, name: data, shape: 1, 3, 640, 640, format: B, C, H, W, where:

  • B - batch size

  • C - channel

  • H - height

  • W - width

Channel order is BGR. Mean values: [104.0, 117.0, 123.0].

Converted model:

Image, name: data, shape: 1, 3, 640, 640, format: B, C, H, W, where:

  • B - batch size

  • C - channel

  • H - height

  • W - width

Channel order is BGR.

Output

Original model

Model outputs are floating points tensors:

  1. name: face_rpn_cls_prob, shape: 1, 16800, 2, format: B, A*C, 2, represents detection scores for 2 classes: background and face.

  2. name: face_rpn_bbox_pred, shape: 1, 16800, 4, format: B, A*C, 4, represents detection box deltas.

  3. name: face_rpn_landmark_pred, shape: 1, 16800, 10, format: B, A*C, 10, represents facial landmarks.

For each output format:

  • B - batch size

  • A - number of anchors

  • C - sum of products of dimensions for each stride, C = H32 * W32 + H16 * W16 + H8 * W8

  • H - feature height with the corresponding stride

  • W - feature width with the corresponding stride

Detection box deltas have format [dx, dy, dh, dw], where:

  • (dx, dy) - regression for center of bounding box

  • (dh, dw) - regression by height and width of bounding box

Facial landmarks have format [x1, y1, x2, y2, x3, y3, x4, y4, x5, y5], where:

  • (x1, y1) - coordinates of left eye

  • (x2, y2) - coordinates of rights eye

  • (x3, y3) - coordinates of nose

  • (x4, y4) - coordinates of left mouth corner

  • (x5, y5) - coordinates of right mouth corner

Converted model

The converted model has the same outputs as the original model.

Download a Model and Convert it into OpenVINO™ IR Format

You can download models and if necessary convert them into OpenVINO™ IR format using the Model Downloader and other automation tools as shown in the examples below.

An example of using the Model Downloader:

omz_downloader --name <model_name>

An example of using the Model Converter:

omz_converter --name <model_name>

Demo usage

The model can be used in the following demos provided by the Open Model Zoo to show its capabilities: