SeasonVision

SeasonVision is a .NET computer vision library built on top of ONNX Runtime. It provides lightweight helpers for common inference scenarios such as image classification, object detection, face detection, face landmarks, face emotion recognition, face attribute analysis, and instance segmentation.

• Repository: SeasonRealms/SeasonVision

Features

• Image classification with MobileNet and Resnet
• Face detection with Ultraface
• Face landmark detection with FaceLandmark
• Face emotion recognition with FaceEmotion
• Face attribute analysis with FaceAttributes
• Object detection with FasterRcnn
• Instance segmentation with MaskRcnn
• Shared image preprocessing utilities for RGBA image buffers

Target Framework

• net10.0

Installation

Add the package reference once a NuGet package is published:

<PackageReference Include="SeasonVision" Version="0.1.0" />

For local development, reference the project directly from your solution.

Model Inputs

SeasonVision does not embed model assets. You are expected to provide compatible ONNX model files when calling the APIs.

The current source includes helpers for models similar to:

• UltraFace face detection
• FER+ emotion classification
• PIPNet face landmark detection
• OpenVINO age-gender-recognition-retail-0013
• Faster R-CNN object detection
• Mask R-CNN instance segmentation
• MobileNet v2 image classification
• ResNet50 image classification

Reference models used by the current helpers:

Local File Name	Upstream Project	Official Weights Page	License	Notes
Age-gender-recognition-retail-0013.onnx	OpenVINO Open Model Zoo	age-gender-recognition-retail-0013	Apache-2.0	Intel retail age/gender model from Open Model Zoo.
Emotion-ferplus-8.onnx	ONNX Model Zoo / FERPlus	emotion-ferplus-8	MIT	The model card body says MIT; the Hugging Face metadata tag also shows apache-2.0, so keep that mismatch in mind when redistributing.
Mobilenetv2-12.onnx	ONNX Model Zoo	mobilenetv2-12	Apache-2.0	Image classification model used by MobileNet.
Resnet50-v1-12-qdq.onnx	ONNX Model Zoo	resnet	Apache-2.0	Quantized ResNet50 variant used by Resnet.
Ultraface_version-RFB-320.onnx	Ultra-Light-Fast-Generic-Face-Detector-1MB	UltraFace	MIT	version-RFB-320 face detector export.
Pipnet_r18_wflw_98.onnx	PIPNet ONNX	pipnet-onnx	MIT	98-point WFLW face landmark model.
FasterRCNN-12-qdq.onnx	ONNX Model Zoo	faster-rcnn	MIT	Quantized Faster R-CNN R-50-FPN model.
MaskRCNN-12-qdq.onnx	ONNX Model Zoo	mask-rcnn	MIT	Quantized Mask R-CNN R-50-FPN model.

API Overview

Most APIs accept:

• model: path to the ONNX model file
• imageData: RGBA byte buffer
• width: image width
• height: image height

Representative entry points:

• Ultraface.Detect(...)
• FaceLandmark.Detect(...)
• FaceEmotion.Detect(...)
• FaceAttributes.Detect(...)
• FasterRcnn.Detect(...)
• MaskRcnn.Detect(...)
• MobileNet.Detect(...)
• Resnet.Detect(...)

Packaging

The project file is prepared for open-source distribution and NuGet packaging with:

• package metadata
• XML documentation generation
• repository metadata
• embedded package README
• symbol package generation with snupkg

To create a package locally:

dotnet pack SeasonVision/SeasonVision.csproj -c Release

Notes

• The library currently uses Microsoft.ML.OnnxRuntime.Managed.
• Some models may require additional runtime/provider configuration depending on your deployment target.
• Annotated image outputs are returned as RGBA byte arrays.

License

This project is released under the MIT License.