NanoDet is a FCOS (Focal Collection of Samples) style one-stage anchor-free object detection model that employs Generalized Focal Loss for both classification and regression. NanoDet-Plus introduces a novel label assignment strategy with an Assign Guidance Module (AGM) and a Dynamic Soft Label Assigner (DSLA). These innovations enhance the model's ability to learn optimal label assignments, resulting in a 7 mAP performance increase on the COCO dataset. Furthermore, a light feature pyramid called Ghost-PAN is incorporated to improve multi-layer feature fusion. This approach streamlines the network architecture and improves accuracy while maintaining computational efficiency.
NanoDet-Plus distinguishes itself through its lightweight architecture, achieving speeds of up to 97 FPS on mobile ARM CPUs while maintaining high accuracy. The introduction of AGM and DSLA significantly improves label assignment, boosting mAP. Its support for multiple backends like ncnn, MNN, and OpenVINO, along with clear and accessible demo implementations (Android, C++, OpenVINO, and Web), makes it easy to integrate and deploy. The provided documentation and comprehensive model zoo further enhance usability.
- Lightweight Design: Model sizes are as low as 980KB (INT8) or 1.8MB (FP16).
- High-Speed Inference: Achieves up to 97 FPS on mobile ARM CPUs.
- Improved Accuracy: Gains 7 mAP on COCO compared to the original NanoDet.
- Flexible Backends: Supports ncnn, MNN, OpenVINO, and PyTorch inference backends.
- Easy Deployment: Provides pre-trained weights and demo code for various platforms (Android, C++, Web).
- Extensive Model Zoo: Supports multiple backbones (ShuffleNet, GhostNet, etc.) with varying sizes and performance trade-offs.
- Extensible Configurations: Offers detailed configuration files (YAML) for easy customization and model tuning.
The project is actively maintained with recent updates focusing on PyTorch Lightning compatibility and performance improvements. A detailed release log is provided to track changes and improvements. The presence of a growing community and comprehensive documentation indicates a healthy development and support ecosystem. The inclusion of a model zoo and demo code further eases adoption and experimentation. The repository is not archived and has recent commits.
NanoDet-Plus benefits developers seeking a high-performance and accurate object detection solution, particularly for resource-constrained environments like mobile devices. It simplifies deployment through backend support and offers a balance of speed and accuracy suitable for real-time applications. Compared to other object detection models, NanoDet-Plus distinguishes itself with its lightweight design, high inference speed, and ease of use, presenting a valuable option for applications requiring both efficiency and accuracy.