Wanli
commited on
Commit
·
576ce22
1
Parent(s):
549df9a
Update hadnpose preprocess (#141)
Browse files* update preprocess of handpose estimation to handle some conner cases
* back PALM_BOX_ENLARGE_FACTOR to default
- mp_handpose.py +64 -36
mp_handpose.py
CHANGED
@@ -13,6 +13,8 @@ class MPHandPose:
|
|
13 |
self.PALM_LANDMARK_IDS = [0, 5, 9, 13, 17, 1, 2]
|
14 |
self.PALM_LANDMARKS_INDEX_OF_PALM_BASE = 0
|
15 |
self.PALM_LANDMARKS_INDEX_OF_MIDDLE_FINGER_BASE = 2
|
|
|
|
|
16 |
self.PALM_BOX_SHIFT_VECTOR = [0, -0.4]
|
17 |
self.PALM_BOX_ENLARGE_FACTOR = 3
|
18 |
self.HAND_BOX_SHIFT_VECTOR = [0, -0.1]
|
@@ -34,6 +36,48 @@ class MPHandPose:
|
|
34 |
self.target_id = targetId
|
35 |
self.model.setPreferableTarget(self.target_id)
|
36 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
def _preprocess(self, image, palm):
|
38 |
'''
|
39 |
Rotate input for inference.
|
@@ -43,14 +87,22 @@ class MPHandPose:
|
|
43 |
palm_landmarks - 7 landmarks (5 finger base points, 2 palm base points) of shape [7, 2]
|
44 |
Returns:
|
45 |
rotated_hand - rotated hand image for inference
|
|
|
|
|
46 |
rotation_matrix - matrix for rotation and de-rotation
|
|
|
47 |
'''
|
48 |
-
#
|
49 |
-
#
|
50 |
palm_bbox = palm[0:4].reshape(2, 2)
|
51 |
-
|
52 |
image = cv.cvtColor(image, cv.COLOR_BGR2RGB)
|
|
|
53 |
|
|
|
|
|
|
|
|
|
54 |
p1 = palm_landmarks[self.PALM_LANDMARKS_INDEX_OF_PALM_BASE]
|
55 |
p2 = palm_landmarks[self.PALM_LANDMARKS_INDEX_OF_MIDDLE_FINGER_BASE]
|
56 |
radians = np.pi / 2 - np.arctan2(-(p2[1] - p1[1]), p2[0] - p1[0])
|
@@ -72,49 +124,25 @@ class MPHandPose:
|
|
72 |
np.amin(rotated_palm_landmarks, axis=1),
|
73 |
np.amax(rotated_palm_landmarks, axis=1)]) # [top-left, bottom-right]
|
74 |
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
rotated_palm_bbox = rotated_palm_bbox + shift_vector
|
79 |
-
# squarify bounding boxx
|
80 |
-
center_rotated_plam_bbox = np.sum(rotated_palm_bbox, axis=0) / 2
|
81 |
-
wh_rotated_palm_bbox = rotated_palm_bbox[1] - rotated_palm_bbox[0]
|
82 |
-
new_half_size = np.amax(wh_rotated_palm_bbox) / 2
|
83 |
-
rotated_palm_bbox = np.array([
|
84 |
-
center_rotated_plam_bbox - new_half_size,
|
85 |
-
center_rotated_plam_bbox + new_half_size])
|
86 |
-
|
87 |
-
# enlarge bounding box
|
88 |
-
center_rotated_plam_bbox = np.sum(rotated_palm_bbox, axis=0) / 2
|
89 |
-
wh_rotated_palm_bbox = rotated_palm_bbox[1] - rotated_palm_bbox[0]
|
90 |
-
new_half_size = wh_rotated_palm_bbox * self.PALM_BOX_ENLARGE_FACTOR / 2
|
91 |
-
rotated_palm_bbox = np.array([
|
92 |
-
center_rotated_plam_bbox - new_half_size,
|
93 |
-
center_rotated_plam_bbox + new_half_size])
|
94 |
-
|
95 |
-
# Crop and resize the rotated image by the bounding box
|
96 |
-
[[x1, y1], [x2, y2]] = rotated_palm_bbox.astype(np.int32)
|
97 |
-
diff = np.maximum([-x1, -y1, x2 - rotated_image.shape[1], y2 - rotated_image.shape[0]], 0)
|
98 |
-
[x1, y1, x2, y2] = [x1, y1, x2, y2] + diff
|
99 |
-
crop = rotated_image[y1:y2, x1:x2, :]
|
100 |
-
crop = cv.copyMakeBorder(crop, diff[1], diff[3], diff[0], diff[2], cv.BORDER_CONSTANT, value=(0, 0, 0))
|
101 |
-
blob = cv.resize(crop, dsize=self.input_size, interpolation=cv.INTER_AREA).astype(np.float32) / 255.0
|
102 |
|
103 |
-
return blob[np.newaxis, :, :, :], rotated_palm_bbox, angle, rotation_matrix
|
104 |
|
105 |
def infer(self, image, palm):
|
106 |
# Preprocess
|
107 |
-
input_blob, rotated_palm_bbox, angle, rotation_matrix = self._preprocess(image, palm)
|
108 |
|
109 |
# Forward
|
110 |
self.model.setInput(input_blob)
|
111 |
output_blob = self.model.forward(self.model.getUnconnectedOutLayersNames())
|
112 |
|
113 |
# Postprocess
|
114 |
-
results = self._postprocess(output_blob, rotated_palm_bbox, angle, rotation_matrix)
|
115 |
return results # [bbox_coords, landmarks_coords, conf]
|
116 |
|
117 |
-
def _postprocess(self, blob, rotated_palm_bbox, angle, rotation_matrix):
|
118 |
landmarks, conf, handedness, landmarks_word = blob
|
119 |
|
120 |
conf = conf[0][0]
|
@@ -127,7 +155,7 @@ class MPHandPose:
|
|
127 |
# transform coords back to the input coords
|
128 |
wh_rotated_palm_bbox = rotated_palm_bbox[1] - rotated_palm_bbox[0]
|
129 |
scale_factor = wh_rotated_palm_bbox / self.input_size
|
130 |
-
landmarks[:, :2] = (landmarks[:, :2] - self.input_size / 2) * scale_factor
|
131 |
landmarks[:, 2] = landmarks[:, 2] * max(scale_factor) # depth scaling
|
132 |
coords_rotation_matrix = cv.getRotationMatrix2D((0, 0), angle, 1.0)
|
133 |
rotated_landmarks = np.dot(landmarks[:, :2], coords_rotation_matrix[:, :2])
|
@@ -149,7 +177,7 @@ class MPHandPose:
|
|
149 |
original_center = np.array([
|
150 |
np.dot(center, inverse_rotation_matrix[0]),
|
151 |
np.dot(center, inverse_rotation_matrix[1])])
|
152 |
-
landmarks[:, :2] = rotated_landmarks[:, :2] + original_center
|
153 |
|
154 |
# get bounding box from rotated_landmarks
|
155 |
bbox = np.array([
|
|
|
13 |
self.PALM_LANDMARK_IDS = [0, 5, 9, 13, 17, 1, 2]
|
14 |
self.PALM_LANDMARKS_INDEX_OF_PALM_BASE = 0
|
15 |
self.PALM_LANDMARKS_INDEX_OF_MIDDLE_FINGER_BASE = 2
|
16 |
+
self.PALM_BOX_PRE_SHIFT_VECTOR = [0, 0]
|
17 |
+
self.PALM_BOX_PRE_ENLARGE_FACTOR = 4
|
18 |
self.PALM_BOX_SHIFT_VECTOR = [0, -0.4]
|
19 |
self.PALM_BOX_ENLARGE_FACTOR = 3
|
20 |
self.HAND_BOX_SHIFT_VECTOR = [0, -0.1]
|
|
|
36 |
self.target_id = targetId
|
37 |
self.model.setPreferableTarget(self.target_id)
|
38 |
|
39 |
+
def _cropAndPadFromPalm(self, image, palm_bbox, for_rotation = False):
|
40 |
+
# shift bounding box
|
41 |
+
wh_palm_bbox = palm_bbox[1] - palm_bbox[0]
|
42 |
+
if for_rotation:
|
43 |
+
shift_vector = self.PALM_BOX_PRE_SHIFT_VECTOR
|
44 |
+
else:
|
45 |
+
shift_vector = self.PALM_BOX_SHIFT_VECTOR
|
46 |
+
shift_vector = shift_vector * wh_palm_bbox
|
47 |
+
palm_bbox = palm_bbox + shift_vector
|
48 |
+
# enlarge bounding box
|
49 |
+
center_palm_bbox = np.sum(palm_bbox, axis=0) / 2
|
50 |
+
wh_palm_bbox = palm_bbox[1] - palm_bbox[0]
|
51 |
+
if for_rotation:
|
52 |
+
enlarge_scale = self.PALM_BOX_PRE_ENLARGE_FACTOR
|
53 |
+
else:
|
54 |
+
enlarge_scale = self.PALM_BOX_ENLARGE_FACTOR
|
55 |
+
new_half_size = wh_palm_bbox * enlarge_scale / 2
|
56 |
+
palm_bbox = np.array([
|
57 |
+
center_palm_bbox - new_half_size,
|
58 |
+
center_palm_bbox + new_half_size])
|
59 |
+
palm_bbox = palm_bbox.astype(np.int32)
|
60 |
+
palm_bbox[:][0] = np.clip(palm_bbox[:][0], 0, image.shape[0])
|
61 |
+
palm_bbox[:][1] = np.clip(palm_bbox[:][1], 0, image.shape[1])
|
62 |
+
# crop to the size of interest
|
63 |
+
image = image[palm_bbox[0][1]:palm_bbox[1][1], palm_bbox[0][0]:palm_bbox[1][0], :]
|
64 |
+
# pad to ensure conner pixels won't be cropped
|
65 |
+
if for_rotation:
|
66 |
+
side_len = np.linalg.norm(image.shape[:2])
|
67 |
+
else:
|
68 |
+
side_len = max(image.shape[:2])
|
69 |
+
|
70 |
+
side_len = int(side_len)
|
71 |
+
pad_h = side_len - image.shape[0]
|
72 |
+
pad_w = side_len - image.shape[1]
|
73 |
+
left = pad_w // 2
|
74 |
+
top = pad_h // 2
|
75 |
+
right = pad_w - left
|
76 |
+
bottom = pad_h - top
|
77 |
+
image = cv.copyMakeBorder(image, top, bottom, left, right, cv.BORDER_CONSTANT, None, (0, 0, 0))
|
78 |
+
bias = palm_bbox[0] - [left, top]
|
79 |
+
return image, palm_bbox, bias
|
80 |
+
|
81 |
def _preprocess(self, image, palm):
|
82 |
'''
|
83 |
Rotate input for inference.
|
|
|
87 |
palm_landmarks - 7 landmarks (5 finger base points, 2 palm base points) of shape [7, 2]
|
88 |
Returns:
|
89 |
rotated_hand - rotated hand image for inference
|
90 |
+
rotate_palm_bbox - palm box of interest range
|
91 |
+
angle - rotate angle for hand
|
92 |
rotation_matrix - matrix for rotation and de-rotation
|
93 |
+
pad_bias - pad pixels of interest range
|
94 |
'''
|
95 |
+
# crop and pad image to interest range
|
96 |
+
pad_bias = np.array([0, 0], dtype=np.int32) # left, top
|
97 |
palm_bbox = palm[0:4].reshape(2, 2)
|
98 |
+
image, palm_bbox, bias = self._cropAndPadFromPalm(image, palm_bbox, True)
|
99 |
image = cv.cvtColor(image, cv.COLOR_BGR2RGB)
|
100 |
+
pad_bias += bias
|
101 |
|
102 |
+
# Rotate input to have vertically oriented hand image
|
103 |
+
# compute rotation
|
104 |
+
palm_bbox -= pad_bias
|
105 |
+
palm_landmarks = palm[4:18].reshape(7, 2) - pad_bias
|
106 |
p1 = palm_landmarks[self.PALM_LANDMARKS_INDEX_OF_PALM_BASE]
|
107 |
p2 = palm_landmarks[self.PALM_LANDMARKS_INDEX_OF_MIDDLE_FINGER_BASE]
|
108 |
radians = np.pi / 2 - np.arctan2(-(p2[1] - p1[1]), p2[0] - p1[0])
|
|
|
124 |
np.amin(rotated_palm_landmarks, axis=1),
|
125 |
np.amax(rotated_palm_landmarks, axis=1)]) # [top-left, bottom-right]
|
126 |
|
127 |
+
crop, rotated_palm_bbox, _ = self._cropAndPadFromPalm(rotated_image, rotated_palm_bbox)
|
128 |
+
blob = cv.resize(crop, dsize=self.input_size, interpolation=cv.INTER_AREA).astype(np.float32)
|
129 |
+
blob = blob / 255.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
130 |
|
131 |
+
return blob[np.newaxis, :, :, :], rotated_palm_bbox, angle, rotation_matrix, pad_bias
|
132 |
|
133 |
def infer(self, image, palm):
|
134 |
# Preprocess
|
135 |
+
input_blob, rotated_palm_bbox, angle, rotation_matrix, pad_bias = self._preprocess(image, palm)
|
136 |
|
137 |
# Forward
|
138 |
self.model.setInput(input_blob)
|
139 |
output_blob = self.model.forward(self.model.getUnconnectedOutLayersNames())
|
140 |
|
141 |
# Postprocess
|
142 |
+
results = self._postprocess(output_blob, rotated_palm_bbox, angle, rotation_matrix, pad_bias)
|
143 |
return results # [bbox_coords, landmarks_coords, conf]
|
144 |
|
145 |
+
def _postprocess(self, blob, rotated_palm_bbox, angle, rotation_matrix, pad_bias):
|
146 |
landmarks, conf, handedness, landmarks_word = blob
|
147 |
|
148 |
conf = conf[0][0]
|
|
|
155 |
# transform coords back to the input coords
|
156 |
wh_rotated_palm_bbox = rotated_palm_bbox[1] - rotated_palm_bbox[0]
|
157 |
scale_factor = wh_rotated_palm_bbox / self.input_size
|
158 |
+
landmarks[:, :2] = (landmarks[:, :2] - self.input_size / 2) * max(scale_factor)
|
159 |
landmarks[:, 2] = landmarks[:, 2] * max(scale_factor) # depth scaling
|
160 |
coords_rotation_matrix = cv.getRotationMatrix2D((0, 0), angle, 1.0)
|
161 |
rotated_landmarks = np.dot(landmarks[:, :2], coords_rotation_matrix[:, :2])
|
|
|
177 |
original_center = np.array([
|
178 |
np.dot(center, inverse_rotation_matrix[0]),
|
179 |
np.dot(center, inverse_rotation_matrix[1])])
|
180 |
+
landmarks[:, :2] = rotated_landmarks[:, :2] + original_center + pad_bias
|
181 |
|
182 |
# get bounding box from rotated_landmarks
|
183 |
bbox = np.array([
|