Mirror of forum.nim-lang.org

12671 :: Face detection?

• angus (orginal) [2025-02-03T03:14:59+01:00] view original

  Hi guys,

  Do you know of any face detection library in Nim? It's a little difficult to find any, with Nvidia naming their new thing NIM...

  Anyway, do you know of any? Preferably not bindings to OpenCV, since I'd like not to have to distribute external libraries with the binary.

  Thanks.

---

• janAkali (orginal) [2025-02-03T06:44:33+01:00] view original

  Most libraries I could find (C, Python, but not Nim unfortunately) are built on top of OpenCV.

  But I also found this little gem: https://github.com/nenadmarkus/pico

  Written in C, without dependencies. Should be very easy to write bindings for.

---

• Araq (orginal) [2025-02-03T09:14:43+01:00] view original

  Written in C, without dependencies. Should be very easy to write bindings for.

  Better use the library which produces best results rather than what is "easiest to wrap at first glance"... Interfacing with Python works too and we have great packages for it in the Nimbleverse.

---

• mig (orginal) [2025-02-03T22:31:59+01:00] view original

  Better use the library which produces best results

  Agreed. This method is from 2013 using old CV techniques.

  You can look at https://paperswithcode.com/task/face-detection for the SOTA techniques - though some of these seem to be using old model architectures (e.g. ResNet, which is computationally cheaper than a stronger architecture such as ViT).

  Here are the paths you could take:

  • (1) Use an existing model
    • (a) Convert an existing PyTorch model to ExecuTorch, ONNX or another standalone run-time (e.g. ggml). See docs:
      • https://pytorch.org/executorch/stable/index.html
      • https://onnxruntime.ai/docs/tutorials/mobile/pose-detection.html
    • Existing APIs (I haven't used these / unsure about quality):
      • inspireface's C/C++ API, which contains RetinaFace but this seems to require OpenCV.
      • https://github.com/DefTruth/lite.ai.toolkit, haven't dug into, not sure on dependencies
  • (2) Train your own model
    • Use paperswithcode as a starting point for datasets (WIDER face seems to be the most common).
    • Then the question is how to train and deploy/integrate your model:
      • Use PyTorch and see 1.a
      • DIY from scratch, use:
        • ggml (wrapper required), or
        • Arraymancer (is this still maintained @mratsim ?), or
        • Roll your own library

  You might run into issues exporting PyTorch models, e.g. see https://discuss.pytorch.org/t/pretrained-model-for-face-detection/62362/7

  Here are some other related projects:

  • mmpose
  • detectron2
  • https://github.com/can-lehmann/exprgrad
  • DarkNet (CNNs/CUDA only) I believe this is no longer maintained due to author leaving CV (see here)

  Side note: this list was a pain to format in RST. Numbered lists seem cause broken rendering, unless I'm doing something wrong.

---

• darkestpigeon (orginal) [2025-02-07T21:44:01+01:00] view original

  For face detection you'd need to do some image manipulation and neural net inferencing. Both are (roughly) in the realm of number crunching, so I'd advise to avoid doing them from scratch (for performance and keeping your sanity reasons). I'd go for OpenCV + OpenVINO (if running of CPU) or TensorRT (if running on nvidia GPU). Wrapping those isn't too bad. Also, as far as I know, both OpenCV and OpenVINO can be built from source and linked statically, but this will probably take some time to figure out.

---