autolane/rfdetr-alpr

RF-DETR License Plate Detector

A fine-tuned RF-DETR Medium model for license plate detection, optimized for edge deployment.

Model Details

• Base Model: RF-DETR Medium
• Task: License plate detection (single class)
• Input Resolution: 576x576
• Training Framework: PyTorch

Available Formats

Deployment Paths

• NVIDIA GPU: Use TensorRT engines (.engine) for fastest inference
• Edge/ARM (i.MX8M Plus, i.MX93): Use INT8 ONNX with ONNX Runtime or convert to TFLite

TensorRT Engine Details

• TensorRT Version: 10.14.1
• Target GPU: NVIDIA GB10 (Compute Capability 12.1)
• Input Shape: 1x3x576x576 (fixed batch size)
• Precision: FP16 / INT8

Usage

ONNX Inference

import onnxruntime as ort
import numpy as np

session = ort.InferenceSession("rfdetr_alpr_optimized.onnx")
# Input: (1, 3, 576, 576) normalized to [0, 1]
outputs = session.run(None, {"images": input_tensor})
boxes, scores = outputs[0], outputs[1]

TensorRT Inference

import tensorrt as trt
import pycuda.driver as cuda

# Load engine
with open("license_plate_detector_int8.engine", "rb") as f:
    engine = trt.Runtime(trt.Logger()).deserialize_cuda_engine(f.read())

Edge Inference (INT8 ONNX)

For ARM/edge devices without NVIDIA GPU:

import onnxruntime as ort

# Use INT8 quantized model for edge deployment
session = ort.InferenceSession(
    "rfdetr_alpr_int8.onnx",
    providers=['CPUExecutionProvider']  # or platform-specific NPU provider
)

# Input: (1, 3, 576, 576) with ImageNet normalization
outputs = session.run(None, {"images": input_tensor})
boxes, scores = outputs[0], outputs[1]

License

Apache 2.0 (same as RF-DETR)