Lite-HRNet | Luxonis

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

0 Likes

Model Details

Model Description

Lite-HRNet is an efficient human pose estimator from the HRNet (high-resolution network) model family. It detects a single human pose with 17 body keypoints (nose, eyes, ears, shoulders, elbows, wrists, hips, knees, and ankles). For optimal performance, it is advisable to couple the model with a human pose detector and take as input human pose cropouts. We implement here the Lite-HRNet-18 and Lite-HRNet-30 variant of the model.

Developed by: Changqian Yu et al.
Shared by:
Model type: Computer Vision
License:
Resources for more information:

Photo by Guy Kawasaki from

Training Details

Training Data

The model was trained on dataset consisting of about 50K real-world images with 150K human pose instances.

Testing Details

Metrics

The model performance was evaluated on dataset calculating various Average Precision (AP) and Average Recall (AR) metrics. See the for more details.

Model	Input Size	AP	AP50	AP75	APM	APL	AR
Lite-HRNet-18	256x192	64.8	86.7	73.0	62.1	70.5	71.2
Lite-HRNet-18	384x288	67.6	87.8	75.0	64.5	73.7	73.7
Lite-HRNet-30	256x192	67.2	88.0	75.0	64.3	73.1	73.3
Lite-HRNet-30	384x288	70.4	88.7	77.7	67.5	76.3	76.2

Technical Specifications

Input/Output Details

Input:
- Name: image
  - Info: NCHW BGR image
Output:
- Name: heatmaps
  - Info: 2D grids where each grid corresponds to a particular keypoint

Model Architecture

The model integrates HRNet's architecture with ShuffleNet's shuffle block node. Additionally, it replaces the expensive pointwise convolution in the shuffle block with "conditional channel weighting". For more details, refer to the .

Throughput

Model variant: lite-hrnet:30-coco-288x384

• Input shape: [1, 3, 384, 288] • Output shape: [1, 17, 96, 72]

• Params (M): 1.739 • GFLOPs: 1.014

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	7.41	N/A
RVC4	FP16	31.66	3.64

Model variant: lite-hrnet:18-coco-192x256

• Input shape: [1, 3, 256, 192] • Output shape: [1, 17, 64, 48]

• Params (M): 1.115 • GFLOPs: 0.281

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	20.09	N/A
RVC4	FP16	119.70	2.85

Model variant: lite-hrnet:30-coco-192x256

• Input shape: [1, 3, 256, 192] • Output shape: [1, 17, 64, 48]

• Params (M): 1.739 • GFLOPs: 0.451

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	11.96	N/A
RVC4	FP16	72.79	3.10

Model variant: lite-hrnet:18-coco-288x384

• Input shape: [1, 3, 384, 288] • Output shape: [1, 17, 96, 72]

• Params (M): 1.115 • GFLOPs: 0.633

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	12.39	N/A
RVC4	FP16	53.08	3.89

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

* Parameters and FLOPs are obtained from the package.

Quantization

RVC4 models were not quantized to int8 due to issues with sigmoid quantization.

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/lite-hrnet:18-coco-288x384"
)

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

Inspect model head(s):

HRNetParser that outputs message (detected body skeleton keypoints).

Get parsed output(s):

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

Example

You can quickly run the model using our example.

The example demonstrates how to build a 2-stage DepthAI pipeline for human pose estimation. The pipeline consists of a pose detection model and a pose estimation model. It automatically downloads the models, 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/lite-hrnet:18-coco-192x256

Human pose estimation model.
License	Apache 2.0 Commercial use
Downloads	1105
Tasks	Keypoint Detection
Model Types	ONNX

Model Variants

Name	Version	Available For	Created At	Deploy
		RVC2, RVC4	8 months ago
		RVC2, RVC4	8 months ago
		RVC2, RVC4	8 months ago
		RVC2, RVC4	8 months ago