X

Model Details

Model Description

This model is based on predefined (light variant) from LuxonisTrain and has been custom-trained to detect people in thermal images that are "cast" into 3-channel grayscale formats. It is specifically tailored for use with OAK Thermal devices (). This design ensures optimized performance for real-world applications in thermal imaging.

Developed by: Luxonis
Shared by: Luxonis
Model type: Computer Vision
License: Apache 2.0
Resources for more information:

Training Details

Training Data

, the , and synthetic datasets were used to train the model. Annotations were filtered by size because smaller objects are barely visible at the camera’s resolution. The dataset consists of 1.5K FLIR images, 1.5K Roboflow images, and 10K synthetic thermal images.

Testing Details

Metrics

The model was trained at a resolution of 384x512. On the testing set, which comprised 10% of the dataset, the model achieved a mean Average Precision (mAP) of 88%.

Technical Specifications

Input/Output Details

Input:
- Name: image
  - Info: NCHW RGB format with images normalized to a range of 0-1.
Output:
- Name: multiple (see NN archive)
  - Info: Unprocessed outputs of a multitude of detections

Model Architecture

Backbone: EfficientRep backbone
Neck: Rep-PAN neck
Head: Efficient decoupled head that is anchor-free

For more details, see the official documentation.

Throughput

Model variant: thermal-person-detection:256x192

• Input shape: [1, 3, 192, 256] • Output shapes: [[1, 6, 24, 32], [1, 6, 12, 16], [1, 6, 6, 8]]

• Params (M): 4.299 • GFLOPs: 0.671

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	189.47	N/A
RVC4	FP32	704.26	2.97

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

* Parameters and FLOPs are obtained from the package.

Utilization

This model is targeted at usage with . Below, we present the most crucial utilization steps for this particular model.

Install DAIv3 library:

pip install depthai
pip install depthai-nodes

Define thermal node for the camera:

cam = pipeline.create(dai.node.Thermal).build()

The output of the camera is in YUV format whereas the model consumes BGR images so that is why we need to define a HostNode that performs this conversion.

class YUV2BGR(dai.node.ThreadedHostNode):
    def __init__(self):
        super().__init__()
        self.input = self.createInput()
        self.out = self.createOutput()

    def run(self):
        while self.isRunning:
            frame = self.input.get().getCvFrame()
            new_frame = dai.ImgFrame()
            new_frame.setCvFrame(frame, dai.ImgFrame.Type.BGR888p)
            self.out.send(new_frame)

yuv2bgr = pipeline.create(YUV2BGR)
cam.color.link(yuv2bgr.input)

And lastly we define the model:

model_description = dai.NNModelDescription(
    "luxonis/thermal-person-detection:256x192"
)

nn = pipeline.create(dai.node.DetectionNetwork).build(
    yuv2bgr.out, model_description
)

The model is automatically parsed by DAI and it outputs the dai.ImgDetections message (bounding boxes, labels, and scores of the detected objects).

Get model output(s):

while pipeline.isRunning():
    nn_output: dai.ImgDetections = parser_output_queue.get()

Example

You can quickly run the model using our example.

The example demonstrates how to build a 1-stage DepthAI pipeline consisting of an thermal person detection model. 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/thermal-person-detection:256x192

Note: This example is designed specifically for RVC2 OAK Thermal.

Person thermal detection model
License	Apache 2.0 Commercial use
Downloads	430
Tasks	Object Detection
Model Types	ONNX