HyperVision is a lightweight anchor-free object detection model, evolved from the NanoDet-Plus architecture. It delivers real-time detection on mobile and edge devices while maintaining competitive accuracy. The model uses Generalized Focal Loss with a ShuffleNetV2 backbone and Ghost-PAN feature pyramid.
| Type | Anchor-free one-stage detector (FCOS-style) |
| Loss | Generalized Focal Loss (QFL + DFL + GIoU) |
| Backbone | ShuffleNetV2 (configurable) |
| Neck | Ghost-PAN feature pyramid |
| Head | NanoDet-Plus head with AGM & DSLA |
| ncnn | Mobile / Android C++ |
| MNN | Mobile / Embedded C++ |
| OpenVINO | Intel CPU / GPU |
| ONNX | Cross-platform |
| LibTorch | C++ inference |
| PyTorch | Python inference |
ShuffleNetV2, ResNet, MobileNetV2, EfficientNet-Lite, GhostNet, RepVGG, Custom CSPNet, TIMM models
To run inference with HyperVision, you can use the GitHub repository or the Marimo notebook below.
@misc{hypervision,
title={HyperVision: Lightweight anchor-free object detection model},
author={Yethikrishna R},
howpublished={\url{https://github.com/Yethikrishna/hypervision}},
year={2025},
note={Premium edition published by Myndlabs.tech}
}