ccv.js

A port of ccv (a minimalistic computer vision library written in C) to the browser using emscripten.

Demos

Usage

Given some example ccv C code you want to translate:

ccv_dense_matrix_t* image = 0;
ccv_read("someFile.png", &image, CCV_IO_GRAY | CCV_IO_ANY_FILE);
ccv_array_t* rects = ccv_swt_detect_words(image, ccv_swt_default_params);
// Do stuff with rects
ccv_array_free(rects);
ccv_matrix_free(image);

The equivalent javascript is:

var image = new CCV.ccv_dense_matrix_t();
CCV.ccv_read(document.querySelector('#some-img-element'), image, CCV.CCV_IO_GRAY);
var rects = CCV.ccv_swt_detect_words(image, CCV.ccv_swt_default_params);
// console.log(rects.toJS());
rects.delete();
image.delete();

Tips for converting:

ccv_read now takes in a CanvasImageSource (which is either <img>, <video> or <canvas>) or ImageData. Only valid flags are CCV.CCV_IO_GRAY or CCV.CCV_IO_RGB_COLOR.
ccv_write now outputs to either a <img>, <canvas>, <div>, or ImageData. If outputing to a div it will append a new canvas with the contents. Can only be used with matrices with datatype CCV_8U (use the getData() method on ccv_dense_matrix_t otherwise).
Only ImageData version will work for reading/writing if using webworkers.
Raw pointers are wrapped in a shared_ptr with corresponding free function as the deleter.
No destructors so you still need to call delete on those shared_ptrs (including the ones returned to you).
Functions that took in a T* now take in a const shared_ptr<T>&.
Functions that took in a T** now take in a shared_ptr<T>& which will hold the output value.
Functions that took in a T*[] now take in a js array of shared_ptr<T>.
Functions that took in a struct param by value now take in a JS object with the same shape (grep for value_object). CCV.ccv_swt_default_params for example is just a JS object.
ccv_array_t needs to know element type (e.g., ccv_rect_array, ccv_comp_array). There is a fromJS() and toJS() method on them to convert from/to js arrays.

See examples/index.js for source code of the demos (which are mostly ported from the C demos in external/ccv/bin/). The bindings were added by hand so if it's not in the demos it probably doesn't have bindings (but they are easy to add yourself).

Install/Build

You should be able to just drop in the js file found at build/ccv.js(~9MB, 2MB gzipped). There is a smaller file at build/ccv_without_filesystem.js(~2.3MB, 400KB gzipped) but at the cost of removing emscripten filesystem support and the model files required for SCD, ICF, and DPM.

If you want rebuild to include your own trained files or add new bindings:

Install emscripten
Run make: emmake make

The emscripten tutorial you should read is this. The C++ bindings are all in ccv_bindings.cpp. JS helpers used in the bindings is in ccv_pre.js.

TODOs / Limitations

Grep for TODO in the code. If it wasn't demoed, it probably doesn't work (notably, convnet doesn't have bindings since it requires a huge sqlite file). Feel free to contribute!

Other interesting javascript computer vision libraries

Also with emscripten:

OpenCV.js (their readme forgot to link to their demos)
tesseract.js (works well with ccv's SWT)

Native JS:

License

MIT for the bindings and BSD 3-clause for ccv (which can be found in ./external/ccv/COPYING)