无法在 facerec 框架中设置阈值（人脸识别）

Question

我是人脸识别的新手。我正在尝试借助 bytefish facerec 框架进行人脸识别。它工作正常，但结果不是很准确。因此，我想设置阈值。按照他的页面 (https://github.com/bytefish/facerec) 上的建议，我应该能够做到。但是，页面上的解释不是很清楚。所以这就是我正在做的事情。

我的分类器

def predict(self, q):
        distances = []
        for xi in self.X:
            xi = xi.reshape(-1,1)
            d = self.dist_metric(xi, q)
            distances.append(d)
        if len(distances) > len(self.y):
            raise Exception("More distances than classes. Is your distance metric correct?")
        distances = np.asarray(distances)
        # Get the indices in an ascending sort order:
        idx = np.argsort(distances)
        # Sort the labels and distances accordingly:
        sorted_y = self.y[idx]
        sorted_distances = distances[idx]
        # Take only the k first items:
        sorted_y = sorted_y[0:self.k]
        sorted_distances = sorted_distances[0:self.k]
        # Make a histogram of them:
        hist = dict((key,val) for key, val in enumerate(np.bincount(sorted_y)) if val)
        # And get the bin with the maximum frequency:
        predicted_label = max(hist.iteritems(), key=op.itemgetter(1))[0]
        # A classifier should output a list with the label as first item and
        # generic data behind. The k-nearest neighbor classifier outputs the 
        # distance of the k first items. So imagine you have a 1-NN and you
        # want to perform a threshold against it, you should take the first
        # item 
        return [predicted_label, { 'labels' : sorted_y, 'distances' : sorted_distances }]

---

我的模型

 def predict(self, X):
        q = self.feature.extract(X)
        return self.classifier.predict(q)

---

我的 server.py 生成输出

def get_prediction(image_data):
    image = preprocess_image(image_data)
    prediction = model.predict(image)
    predicted_label = prediction[0]
    classifier_output = prediction[1]
    distance = classifier_output['distances'][0]
    #distance = classifier.predict(self, q)
    #distance = 11
    if distance > 10.0:
        return "nonsense"
    else:
        print prediction

所以问题是我无法在这里获得距离。请帮忙

Answer 1

一段时间后，我能够解决问题。阈值应该在分类器文件中完成，而不是在 server.py 中。

解决方案

        distances = []
        for xi in self.X:
            xi = xi.reshape(-1,1)
            d = self.dist_metric(xi, q)
            distances.append(d)
        if len(distances) > len(self.y):
            raise Exception("More distances than classes. Is your distance metric correct?")
        distances = np.asarray(distances)
        # Get the indices in an ascending sort order:
        idx = np.argsort(distances)
        # Sort the labels and distances accordingly:
        sorted_y = self.y[idx]
        sorted_distances = distances[idx]
        # Take only the k first items:
        sorted_y = sorted_y[0:self.k]
        sorted_distances = sorted_distances[0:self.k]
        #sorted_distances = 1134.04873217
        # Make a histogram of them:
        hist = dict((key,val) for key, val in enumerate(np.bincount(sorted_y)) if val)
        # And get the bin with the maximum frequency:
        predicted_label = max(hist.iteritems(), key=op.itemgetter(1))[0]
            # A classifier should output a list with the label as first item and
            # generic data behind. The k-nearest neighbor classifier outputs the 
            #global unknown
        if sorted_distances > 1800 :
                return [predicted_label]
        else:
                return [predicted_label]