If you think it works slow, make sure it is compiled with any of these instructions
> pdlib
>
> PDlib support => enabled
> Pdlib Version => 1.0.2
> DLIB_USE_CUDA => false
> DLIB_USE_BLAS => true
> DLIB_USE_LAPACK => true
> USE_AVX_INSTRUCTIONS => false
> USE_AVX2_INSTRUCTIONS => false
> USE_NEON_INSTRUCTIONS => false
> USE_SSE2_INSTRUCTIONS => true
> USE_SSE4_INSTRUCTIONS => false
This small optimization allows us to improve from:
OLD php: 0.012100038051605 Sec.
New native: 0.0096198790073395 Sec.
(This is already 20 percent faster)
..to:
OLD php: 0.012276937961578 Sec.
New native: 0.0022540950775146 Sec.
What is 80 percent faster than our PHP class.
It not use the 'length' native versions of dlib, due I have to convert the
array to the vector, which will take the same time, and then I will use the
native method.
This change adds support to retrieve 128D face descriptor for a given
landmark. Since now we have full pipeline, README.md has "general usage"
section and integration test is added. Also, return from
FaceLandmarkDetection is changed, so it can be given to FaceRecognition
without changes. All obtained values are crosschecked to match with values
from python versions (however, if num_jitters is > 1 in FaceRecognition,
values don't match between PHP and Python, I suspect it is related to usage
of dlib::rand, but still investigating)..
This change extends existing landmark detection in new ways:
1. Existing logic is hiding HOG model (frontal_face_detector)
underneath and user cannot use other models (CNN model, for example).
2. Bounding box is exposed as additional argument, and user
can define custom bounding box (which is needed, if image used to detect
faces is changed (for example scaled), and we want to crop only face
from original image to feed into shape predictor).
3. This approach is class-based, so no need for multiple loadings of
shape predictor model (only once, in ctor)
This change extends existing landmark detection in new ways:
1. Existing logic is hiding HOG model (frontal_face_detector)
underneath and user cannot use other models (CNN model, for example).
2. Bounding box is exposed as additional argument, and user
can define custom bounding box (which is needed, if image used to detect
faces is changed (for example scaled), and we want to crop only face
from original image to feed into shape predictor).
3. This approach is class-based, so no need for multiple loadings of
shape predictor model (only once, in ctor)
This commit introduces chinese_whispers method without any "helpers" on
PHP side. Users needs to take care of everything (building edges from
128D face chip vector, for example), but this is exposed for people
that need low-level call and want to calculate distances and build edges
in PHP directly.
This commit introduces chinese_whispers method without any "helpers" on
PHP side. Users needs to take care of everything (building edges from
128D face chip vector, for example), but this is exposed for people
that need low-level call and want to calculate distances and build edges
in PHP directly.
CNN face detector (deep learning face detection) is modeled as
PHP class. Currently, it only has constructor (which loads a model)
and detect() method. I tried to make it resamble to Python implementation.
Similar to Python, also added support for upsampling. Returned value
is array with numbers as keys where number of keys is equal to number
of found faced. Each element is associative array that has "top", "left",
"right", "bottom" and "detection_confidence" keys.
All errors are propagated using exceptions.
Added two new error tests. I didn't want to rely on presence of model,
so tests are not having great coverage. If we are OK to download models
in tests, it would allow us to have far better coverage.
What is missing:
* Testing with models (should we test that?)
* detect_mult method (similar to what Python have; however, I consider
that this should not block pushing this change)
Matias De lellis