How to parse the heatmap output for the pose estimation tflite model?

Question

I am starting with the pose estimation tflite model for getting keypoints on humans.

https://www.tensorflow.org/lite/models/pose_estimation/overview

I have started with fitting a single image or a person and invoking the model:

img = cv.imread('photos\standing\\3.jpg')
img = tf.reshape(tf.image.resize(img, [257,257]), [1,257,257,3])
model = tf.lite.Interpreter('models\posenet_mobilenet_v1_100_257x257_multi_kpt_stripped.tflite')
model.allocate_tensors()
input_details = model.get_input_details()
output_details = model.get_output_details()
floating_model = input_details[0]['dtype'] == np.float32
if floating_model:
    img = (np.float32(img) - 127.5) / 127.5
model.set_tensor(input_details[0]['index'], img)
model.invoke()
output_data =  model.get_tensor(output_details[0]['index'])# o()
offset_data = model.get_tensor(output_details[1]['index'])
results = np.squeeze(output_data)
offsets_results = np.squeeze(offset_data)
print("output shape: {}".format(output_data.shape))
np.savez('sample3.npz', results, offsets_results)

but I am struggling with parsing the output correctly to get the coordinates/confidences of each body part. Does anyone have a python example for interpreting this models results? (for example: using them to map keypoints back to the original image)

My code (a snippet from a class which essentially takes the np array directly from the model output):

def get_keypoints(self, data):
        height, width, num_keypoints = data.shape
        keypoints = []
        for keypoint in range(0, num_keypoints):
            maxval = data[0][0][keypoint]
            maxrow = 0
            maxcol = 0
            for row in range(0, width):
                for col in range(0,height):
                    if data[row][col][keypoint] > maxval:
                        maxrow = row
                        maxcol = col
                        maxval = data[row][col][keypoint]
            keypoints.append(KeyPoint(keypoint, maxrow, maxcol, maxval))
            # keypoints = [Keypoint(x,y,z) for x,y,z in ]
        return keypoints
def get_image_coordinates_from_keypoints(self, offsets):
        height, width, depth = (257,257,3)
        # [(x,y,confidence)]
        coords = [{ 'point': k.body_part,
                    'location': (k.x / (width - 1)*width + offsets[k.y][k.x][k.index],
                   k.y / (height - 1)*height + offsets[k.y][k.x][k.index]),
                    'confidence': k.confidence}
                 for k in self.keypoints]
        return coords

after matching the indexes to the parts my output is:

Some of the coordinates here are negative, which can't be correct. Where is my mistake?

Answer 1

import numpy as np

For a pose estimation model which outputs a heatmap and offsets. The desired points can be obtained by:

1. Performing a sigmoid operation on the heatmap:

   scores = sigmoid(heatmaps)

2. Each keypoint of those pose is usually represented by a 2-D matrix, the maximum value in that matrix is related to where the model thinks that point is located in the input image. Use argmax2D to obtain the x and y index of that value in each matrix, the value itself represents the confidence value:

   x,y = np.unravel_index(np.argmax(scores[:,:,keypointindex]), scores[:,:,keypointindex].shape) confidences = scores[x,y,keypointindex]

3. That x,y is used to find the corresponding offset vector for calculating the final location of the keypoint:

   offset_vector = (offsets[y,x,keypointindex], offsets[y,x,num_keypoints+keypointindex])

4. After you have obtained your keypoint coords and offsets you can calculate the final position of the keypoint by using ():

   image_positions = np.add(np.array(heatmap_positions) * output_stride, offset_vectors)

See this for determining how to get the output stride, if you don't already have it. The tflite pose estimation has an output stride of 32.

A function which takes output from that Pose Estimation model and outputs keypoints. Not including KeyPoint class

def get_keypoints(self, heatmaps, offsets, output_stride=32):
        scores = sigmoid(heatmaps)
        num_keypoints = scores.shape[2]
        heatmap_positions = []
        offset_vectors = []
        confidences = []
        for ki in range(0, num_keypoints ):
            x,y = np.unravel_index(np.argmax(scores[:,:,ki]), scores[:,:,ki].shape)
            confidences.append(scores[x,y,ki])
            offset_vector = (offsets[y,x,ki], offsets[y,x,num_keypoints+ki])
            heatmap_positions.append((x,y))
            offset_vectors.append(offset_vector)
        image_positions = np.add(np.array(heatmap_positions) * output_stride, offset_vectors)
        keypoints = [KeyPoint(i, pos, confidences[i]) for i, pos in enumerate(image_positions)]
        return keypoints

Keypoint class:

PARTS = {
    0: 'NOSE',
    1: 'LEFT_EYE',
    2: 'RIGHT_EYE',
    3: 'LEFT_EAR',
    4: 'RIGHT_EAR',
    5: 'LEFT_SHOULDER',
    6: 'RIGHT_SHOULDER',
    7: 'LEFT_ELBOW',
    8: 'RIGHT_ELBOW',
    9: 'LEFT_WRIST',
    10: 'RIGHT_WRIST',
    11: 'LEFT_HIP',
    12: 'RIGHT_HIP',
    13: 'LEFT_KNEE',
    14: 'RIGHT_KNEE',
    15: 'LEFT_ANKLE',
    16: 'RIGHT_ANKLE'
}

class KeyPoint():
    def __init__(self, index, pos, v):
        x, y = pos
        self.x = x
        self.y = y
        self.index = index
        self.body_part = PARTS.get(index)
        self.confidence = v

    def point(self):
        return int(self.y), int(self.x)

    def to_string(self):
        return 'part: {} location: {} confidence: {}'.format(
            self.body_part, (self.x, self.y), self.confidence)