How to preserve YOLO11 + PaddleOCR v5 accuracy when compiling to HEF for Hailo-15L?

We are a startup building a production ANPR camera on Hailo-15L. Our current ANPR stack uses a custom YOLO11l license-plate detector and PaddleOCR v5. These models already perform well on NVIDIA Jetson/GPU, and our main requirement is that moving them to Hailo should not materially reduce detection or OCR accuracy. We are willing to trade some FPS/resources for accuracy if required.

We would appreciate guidance on these specific points:

  1. YOLO11l → HEF: What is the recommended DFC workflow to preserve accuracy as closely as possible to the original PyTorch/ONNX model? In particular, calibration dataset selection/size, INT8 vs higher/mixed precision, QAT, and per-layer precision overrides.

  2. During YOLO11 compilation we had to modify some unsupported graph operations/attention/softmax behavior to make the model compile. What is the recommended Hailo-supported implementation for YOLO11 that preserves the original model numerics/accuracy?

  3. Accuracy validation: What is the recommended way to compare PyTorch/ONNX vs HAR vs HEF accuracy and identify exactly which layers or quantization steps are causing degradation? Is there an official accuracy-analysis/profiling workflow?

  4. PaddleOCR: What exact PaddleOCR version/checkpoint is the supplied Hailo OCR HEF based on? We currently use PaddleOCR v5. Can we compile the latest PaddleOCR v5 recognition model ourselves for Hailo-15L, and what is the recommended export path?

  5. For OCR specifically, what calibration/QAT/mixed-precision strategy is recommended to preserve character recognition accuracy, especially for small, blurred, angled and low-light license plates?

Our priority is production accuracy parity with the original models, rather than simply getting the models to compile and run. If Hailo has recommended recipes, reference models, calibration guidelines, or tools for minimizing the accuracy gap, please share them.

Thanks!

The Hailo Community currently supports our AI Accelerator products only.
For AI Vision Processors questions, please fill out the product inquiry form, and our team will get back to you shortly. Hailo AI Product Inquiry: Connect with Our AI Experts