FastPreProcess

Implemented

1. hwc to chw (slower than numpy.transpose): array(768, 1024, 3) -> array(3, 768, 1024)
2. chw channel normalization: e.g. array(3, 768, 1024) -> (array(3, 768, 1024) - MEAN) / STD
3. hwc to chw with normalization (the fastest): e.g. array(768, 1024, 3) -> (array(3, 768, 1024) - MEAN) / STD
4. hwc to chw with normalization_batched: e.g. array(32, 768, 1024, 3) -> (array(32, 3, 768, 1024) - MEAN) / STD

Performance

• Batch Iteration / second

• MIN / MAX / AVG

• Test Image: (640, 427, 3)

• Test Envirnoment

  • Ubuntu 22.04
  • Intel i9-13900K
  • Python 3.9

• hwc to chw with normalization, hwc to chw with normalization_batched:

Batch Size	Python	Cpp	Cpp (batched)
32	23.67 / 24.43 / 24.11	203.49 / 343.20 / 245.02	207.02 / 266.26 / 228.76

Requirements

1. pybind11
2. numpy
3. opencv

Build

# Edit set(Python3_ROOT_DIR "/path/to/your_python_env") in CMakeLists.txt

mkdir -p build
cmake --build ./build --config Release --target all -j 34 --