defmodule Evision.Zoo.FaceRecognition.SFace do
@moduledoc """
SFace: Sigmoid-Constrained Hypersphere Loss for Robust Face Recognition
"""
@doc """
Default configuration.
"""
@spec default_config :: map()
def default_config do
%{
backend: Evision.Constant.cv_DNN_BACKEND_OPENCV(),
target: Evision.Constant.cv_DNN_TARGET_CPU(),
distance_type: :cosine_similarity,
cosine_threshold: 0.363,
l2_norm_threshold: 1.128,
conf_threshold: 0.9,
nms_threshold: 0.3
}
end
@doc """
Customizable parameters from smart cell.
"""
@spec smartcell_params() :: Evision.Zoo.smartcell_params()
def smartcell_params() do
config = default_config()
[
%{
name: "Face Recognizer",
params: [
%{field: "detector", label: "Face Detector", type: :string, default: "yunet",
is_option: true,
options: [
%{value: "yunet", label: "YuNet"},
%{value: "yunet_quant", label: "YuNet (Quant)"}
]
},
%{field: "distance_type", label: "Distance Type", type: :string, default: "#{config[:distance_type]}",
is_option: true,
options: [
%{value: "cosine_similarity", label: "Cosine Similarity"},
%{value: "l2_norm", label: "L2 Norm"},
]
},
%{field: "cosine_threshold", label: "Cosine Threshold", type: :float, default: config[:cosine_threshold]},
%{field: "l2_norm_threshold", label: "L2-norm Threshold", type: :float, default: config[:l2_norm_threshold]}
]
}
] ++ Evision.Zoo.FaceDetection.YuNet.smartcell_params()
end
@doc """
Initialize model.
##### Positional arguments
- **model**: `String.t()` | `:default_model` | `:quant_model`.
- When `model` is a string, it will be treat as the path to a weight file
and `init/2` will load the model from it.
- When `model` is either `:default_model` or `:quant_model`, `init/2` will
download and load the predefined model.
##### Keyword arguments
- **cache_dir**: `String.t()`.
Path to the cache directory.
Optional. Defaults to `:filename.basedir(:user_cache, "", ...)`
- **backend**: `integer()`.
Specify the backend.
Optional. Defaults to `Evision.Constant.cv_DNN_BACKEND_OPENCV()`.
- **target**: `integer()`.
Specify the target.
Optional. Defaults to `Evision.Constant.cv_DNN_TARGET_CPU()`.
"""
@spec init(binary | :default_model | :quant_model, nil | Keyword.t()) :: {:error, String.t()} | Evision.FaceRecognizerSF.t()
def init(model_path, opts \\ [])
def init(model_type, opts) when model_type in [:default_model, :quant_model] do
{model_url, filename} = model_info(model_type)
cache_dir = opts[:cache_dir]
with {:ok, local_path} <- Evision.Zoo.download(model_url, filename, cache_dir: cache_dir) do
init(local_path, opts)
else
{:error, msg} ->
raise msg
end
end
def init(model_path, opts) when is_binary(model_path) do
config = default_config()
backend = opts[:backend] || config[:backend]
target = opts[:target] || config[:target]
Evision.FaceRecognizerSF.create(
model_path, "",
backend_id: backend,
target_id: target
)
end
@doc """
Get feature for the input face image.
##### Positional arguments
- **self**: `Evision.FaceRecognizerSF.t()`.
An initialized FaceRecognizerSF model.
- **face_image**: `Evision.Mat.maybe_mat_in()`.
Input face image.
- **bbox**: `Evision.Mat.maybe_mat_in()`.
Optional bounding box that specifies the face location in the given image.
Defaults to `nil`.
When `bbox` is not `nil`, `preprocess/3` will crop the face from given image
and use the cropped image as the input.
Otherwise, `face_image` will be set as the input.
"""
def infer(self, face_image, bbox \\ nil) do
inputBlob = preprocess(self, face_image, bbox)
Evision.FaceRecognizerSF.feature(self, inputBlob)
end
@doc """
Preprocessing the input face image.
`infer/3` will call this function automatically.
##### Positional arguments
- **self**: `Evision.FaceRecognizerSF.t()`.
An initialized FaceRecognizerSF model.
- **face_image**: `Evision.Mat.maybe_mat_in()`.
Input face image.
- **bbox**: `Evision.Mat.maybe_mat_in()`.
Optional bounding box that specifies the face location in the given image.
Defaults to `nil`.
When `bbox` is not `nil`, `preprocess/3` will crop the face from given image
and return the cropped image.
Otherwise, `face_image` will be returned.
"""
def preprocess(_self, face_image, nil), do: face_image
def preprocess(self, face_image, bbox) do
Evision.FaceRecognizerSF.alignCrop(self, face_image, bbox)
end
@doc """
Compare two face features.
##### Positional Arguments
- **self**: `Evision.FaceRecognizerSF.t()`
- **face1_feat**: `Evision.Mat.maybe_mat_in()`.
Feature value of face 1.
- **face2_feat**: `Evision.Mat.maybe_mat_in()`.
Feature value of face 2.
##### Keyword Arguments
- **distance_type**: `atom`.
Either `:cosine_similarity` or `:l2_norm`. Defaults to `:cosine_similarity`.
- **cosine_threshold**: `number()`.
Defaults to `0.363`.
- **l2_norm_threshold**: `number()`.
Defaults to `1.128`.
##### Return
A map with three keys.
- **matched**: `boolean()`
`true` if two faces match, `false` otherwise.
- **measure**: `"cosine_score"` | "`l2_norm_distance`"
Distance type.
- **retval**: `number()`
- When `measure == "cosine_score"`, `retval` is the cosine similarity score.
- When `measure == "l2_norm_distance"`, `retval` is the L2 norm distance.
"""
@spec match_feature(Evision.FaceRecognizerSF.t(), Evision.Mat.maybe_mat_in(), Evision.Mat.maybe_mat_in()) :: any()
def match_feature(self=%Evision.FaceRecognizerSF{}, face1_feat, face2_feat, opts \\ []) do
config = default_config()
distance_type = opts[:distance_type] || config[:distance_type]
distance_type =
case distance_type do
:cosine_similarity ->
0
:l2_norm ->
1
_ ->
raise "Unknown distance_type: #{inspect(distance_type)}"
end
result = Evision.FaceRecognizerSF.match(self, face1_feat, face2_feat, dis_type: distance_type)
distance_type = opts[:distance_type] || :cosine_similarity
get_result(distance_type, result, opts)
end
defp get_result(:cosine_similarity, cosine_score, opts) do
default_config = default_config()
cosine_threshold = opts[:cosine_threshold] || default_config[:cosine_threshold]
%{matched: cosine_score >= cosine_threshold, retval: cosine_score, measure: "cosine_score"}
end
defp get_result(:l2_norm, l2_norm_distance, opts) do
default_config = default_config()
l2_norm_threshold = opts[:l2_norm_threshold] || default_config[:l2_norm_threshold]
%{matched: l2_norm_distance < l2_norm_threshold, retval: l2_norm_distance, measure: "l2_norm_distance"}
end
defp get_result(distance_type, _, _) do
raise "Unknown distance_type: #{inspect(distance_type)}"
end
@doc """
Compare two faces.
##### Positional Arguments
- **self**: `Evision.FaceRecognizerSF.t()`
- **original**: `Evision.Mat.maybe_mat_in()`.
Original face.
- **comparison**: `Evision.Mat.maybe_mat_in()`.
Comparison face.
##### Keyword Arguments
- **distance_type**: `atom`.
Either `:cosine_similarity` or `:l2_norm`. Defaults to `:cosine_similarity`.
- **cosine_threshold**: `number()`.
Defaults to `0.363`.
- **l2_norm_threshold**: `number()`.
Defaults to `1.128`.
- **detector_module**: `module()`.
Face detector module. Defaults to `Evision.Zoo.FaceDetection.YuNet`.
- **detector_model**: `String.t()` | `atom()`.
Face detector model name or path to model weights. Defaults to `:default_model`.
- **detector_opts**: `Keyword.t()`.
Face detector initialization options. Defaults to `[]`.
##### Return
A map with three keys.
- **matched**: `boolean()`
`true` if two faces match, `false` otherwise.
- **measure**: `"cosine_score"` | "`l2_norm_distance`"
Distance type.
- **retval**: `number()`
- When `measure == "cosine_score"`, `retval` is the cosine similarity score.
- When `measure == "l2_norm_distance"`, `retval` is the L2 norm distance.
"""
@spec match(Evision.FaceRecognizerSF.t(), Evision.Mat.maybe_mat_in(), Evision.Mat.maybe_mat_in()) :: any()
def match(self=%Evision.FaceRecognizerSF{}, original, comparison, opts \\ []) do
detector_module = opts[:detector_module] || Evision.Zoo.FaceDetection.YuNet
detector_model = opts[:detector_model] || :default_model
detector_opts = opts[:detector_opts] || []
detector = detector_module.init(detector_model, detector_opts)
original_face_bbox = detector_module.infer(detector, original)
comparison_face_bbox = detector_module.infer(detector, comparison)
original_face_feature = Evision.Zoo.FaceRecognition.SFace.infer(self, original, original_face_bbox)
comparison_face_feature = Evision.Zoo.FaceRecognition.SFace.infer(self, comparison, comparison_face_bbox)
match_feature(self, original_face_feature, comparison_face_feature, opts)
end
@doc """
Model URL and filename of predefined model.
"""
@spec model_info(:default_model | :quant_model) :: {String.t(), String.t()}
def model_info(:default_model) do
{
"https://github.com/opencv/opencv_zoo/blob/master/models/face_recognition_sface/face_recognition_sface_2021dec.onnx?raw=true",
"face_recognition_sface_2021dec.onnx"
}
end
def model_info(:quant_model) do
{
"https://github.com/opencv/opencv_zoo/blob/master/models/face_recognition_sface/face_recognition_sface_2021dec-act_int8-wt_int8-quantized.onnx?raw=true",
"face_recognition_sface_2021dec-act_int8-wt_int8-quantized.onnx"
}
end
@doc """
Docs in smart cell.
"""
@spec docs() :: String.t()
def docs do
@moduledoc
end
@doc """
Smart cell tasks.
A list of variants of the current model.
"""
@spec smartcell_tasks() :: Evision.Zoo.smartcell_tasks()
def smartcell_tasks do
[
%{
id: "sface",
label: "SFace",
docs_url: "https://github.com/opencv/opencv_zoo/tree/master/models/face_recognition_sface",
params: smartcell_params(),
docs: docs(),
},
%{
id: "sface_quant",
label: "SFace (quant)",
docs_url: "https://github.com/opencv/opencv_zoo/tree/master/models/face_recognition_sface",
params: smartcell_params(),
docs: docs(),
},
]
end
@doc """
Generate quoted code from smart cell attrs.
"""
@spec to_quoted(map()) :: list()
def to_quoted(attrs) do
{backend, target} = Evision.Zoo.to_quoted_backend_and_target(attrs)
recognizer_opts = [
backend: backend,
target: target,
distance_type: String.to_atom(attrs["distance_type"]),
cosine_threshold: attrs["cosine_threshold"],
l2_norm_threshold: attrs["l2_norm_threshold"]
]
detector_opts = [
backend: backend,
target: target,
nms_threshold: attrs["nms_threshold"],
conf_threshold: attrs["conf_threshold"],
top_k: 5
]
model =
case attrs["variant_id"] do
"sface_quant" ->
:quant_model
_ ->
:default_model
end
{detector_module, detector_model} =
case attrs["detector"] do
"yunet" ->
{Evision.Zoo.FaceDetection.YuNet, :default_model}
"yunet_quant" ->
{Evision.Zoo.FaceDetection.YuNet, :quant_model}
unknown_detector ->
raise "Unknown face detector: #{inspect(unknown_detector)}"
end
[
quote do
recognizer = Evision.Zoo.FaceRecognition.SFace.init(unquote(model), unquote(recognizer_opts))
detector = unquote(detector_module).init(unquote(detector_model), unquote(detector_opts))
end,
quote do
original_input = Kino.Input.image("Original")
comparison_input = Kino.Input.image("Comparison")
form = Kino.Control.form([original: original_input, comparison: comparison_input], submit: "Run")
frame = Kino.Frame.new()
form
|> Kino.Control.stream()
|> Stream.filter(& (&1.data.original != nil) or (&1.data.comparison != nil))
|> Kino.listen(fn %{data: %{original: original_image, comparison: comparison_image}} ->
Kino.Frame.render(frame, Kino.Markdown.new("Running..."))
original_image = Evision.Mat.from_binary(original_image.data, {:u, 8},
original_image.height, original_image.width, 3)
comparison_image = Evision.Mat.from_binary(comparison_image.data, {:u, 8},
comparison_image.height, comparison_image.width, 3)
original_results = unquote(detector_module).infer(detector, original_image)
comparison_results = unquote(detector_module).infer(detector, comparison_image)
case {original_results, comparison_results} do
{%Evision.Mat{}, %Evision.Mat{}} ->
original_bbox = Nx.reverse(Evision.Mat.to_nx(original_results, Nx.BinaryBackend)[0])
comparison_bbox = Nx.reverse(Evision.Mat.to_nx(comparison_results, Nx.BinaryBackend)[0])
original_feature = Evision.Zoo.FaceRecognition.SFace.infer(recognizer, original_image, original_bbox)
comparison_feature = Evision.Zoo.FaceRecognition.SFace.infer(recognizer, comparison_image, comparison_bbox)
%{matched: matched, retval: val, measure: measure} = Evision.Zoo.FaceRecognition.SFace.match_feature(
recognizer, original_feature, comparison_feature)
original_image = Evision.cvtColor(original_image, Evision.Constant.cv_COLOR_RGB2BGR())
comparison_image = Evision.cvtColor(comparison_image, Evision.Constant.cv_COLOR_RGB2BGR())
vis_original = unquote(detector_module).visualize(original_image, original_results[0])
vis_comparison = unquote(detector_module).visualize(comparison_image, comparison_results[0])
vis = [
Kino.Image.new(Evision.imencode(".png", vis_original), :png),
Kino.Image.new(Evision.imencode(".png", vis_comparison), :png)
]
Kino.Frame.render(frame, Kino.Layout.grid(vis, columns: 2))
Kino.Frame.append(frame, Kino.Markdown.new("Result: #{matched}, #{measure}: #{val}"))
{{:error, _}, %Evision.Mat{}} ->
Kino.Frame.render(frame, Kino.Markdown.new("Cannot detect any face in the original image"))
{%Evision.Mat{}, _} ->
Kino.Frame.render(frame, Kino.Markdown.new("Cannot detect any face in the comparison image"))
{_, _} ->
Kino.Frame.render(frame, Kino.Markdown.new("Cannot detect any face in both original and comparison images"))
end
end)
Kino.Layout.grid([form, frame], boxed: true, gap: 16)
end
]
end
end
