SCRFD Person detection

Our new model ZOO works with DepthAI V3. Find out more in our documentation.

0 Likes

Model Details

Model Description

SCRFD was originally developed and trained for face detection tasks but authors also trained it on person detection tasks as it turns out it works pretty well. You can read more about the SCRFD model in a SCRFD model card. Overall it is a lightweight, fast, and accurate person detector. It outputs bounding boxes with appropriate scores and 5 keypoints - the center and corners of the bounding box.

Developed by: InsightFace
Shared by:
Model type: Computer Vision
License:
Resources for more information:

Training Details

Training Data

The model was trained on . It is split into training, validation, and test sets. WiderPerson contains a total of 13,382 images with 399,786 annotations, i.e., 29.87 annotations per image, which means this dataset contains dense pedestrians with various kinds of occlusions.

Testing Details

Metrics

Results for person detection task are not given so we are listing the evaluation results for the same architecture for face detection task.

Evaluation of the model on WIDERFace dataset for all three categories: Easy, Medium and Hard. Results are taken from .

Category	mAP
Easy	93.78
Medium	92.16
Hard	77.87

Technical Specifications

Input/Output Details

Input:
- Name: input
  - Info: NCHW BGR un-normalized image
Output:
- Name: multiple (see NN archive)
  - Info: Classification scores, bounding boxes, and keypoints for a multitude of detections.

Model Architecture

Backbone: ResNet backbone
Neck: Path Aggregation Feature Pyramid Network (PAFPN) neck
Head: Simple head consisting of stacked 3 × 3 convolutional layers

Please consult the for more information on model architecture.

Throughput

Model variant: scrfd-person-detection:25g-640x640

• Input shape: [1, 3, '?', '?'] • Output shapes:

[[6400, 1], [1600, 1], [400, 1], [100, 1], [25, 1], [6400, 4], [1600, 4], [400, 4], [100, 4], [25, 4], [6400, 10], [1600, 10], [400, 10], [100, 10], [25, 10]]

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	27.50	N/A
RVC4	FP32	323.49	3.68

* Benchmarked with , using 2 threads (and the DSP runtime in balanced mode for RVC4).

* Parameters and FLOPs are obtained from the package.

Quantization

RVC4 version of the model was quantized using a custom dataset. This was created by taking 50-image subset of dataset.

Utilization

Models converted for RVC Platforms can be used for inference on OAK devices. DepthAI pipelines are used to define the information flow linking the device, inference model, and the output parser (as defined in model head(s)). Below, we present the most crucial utilization steps for the particular model. Please consult the docs for more information.

Install DAIv3 and depthai-nodes libraries:

pip install depthai
pip install depthai-nodes

Define model:

model_description = dai.NNModelDescription(
    "luxonis/scrfd-person-detection:25g-640x640"
)

nn = pipeline.create(ParsingNeuralNetwork).build(
    <CameraNode>, model_description
)

Inspect model head(s):

SCRFDParser that outputs message (bounding boxes and confidence scores for every detected person).

Get parsed output(s):

while pipeline.isRuning():
    parser_output: ImgDetectionsExtended = parser_output_queue.get()

Example

You can quickly run the model using our script. It automatically downloads the model, creates a DepthAI pipeline, runs the inference, and displays the results using our DepthAI visualizer tool. To try it out, run:

python3 main.py \
    --model luxonis/scrfd-person-detection:25g-640x640

Moreover, you can also check the and examples.

Person detection model.
License	Apache 2.0 Commercial use
Downloads	575
Tasks	Object Detection
Model Types	ONNX

Model Variants

Name	Version	Available For	Created At	Deploy
		RVC2, RVC3, RVC4	10 months ago