def test_non_max_suppression_fast(self):
# 3 boxes which overlap at 1,1 and another 3 boxes at 4,4
boxes = np.array([
[1.0, 1.0, 2, 2],
[1.1, 1.1, 2, 2],
[1.1, 1.1, 1.9, 1.9],
[4.0, 4.0, 2, 2],
[4.1, 4.1, 2, 2],
[4.1, 4.1, 1.9, 1.9]
])
picks = non_max_suppression_fast(boxes)
self.assertEqual(len(picks), 2)
def detect_hand(self, img_norm):
assert -1 <= img_norm.min() and img_norm.max() <= 1,\
"img_norm should be in range [-1, 1]"
assert img_norm.shape == (256, 256, 3),\
"img_norm shape must be (256, 256, 3)"
# predict hand location and 7 initial landmarks
self.interp_palm.set_tensor(self.in_idx, img_norm[None])
self.interp_palm.invoke()
"""
out_reg shape is [number of anchors, 18]
Second dimension 0 - 4 are bounding box offset, width and height: dx, dy, w ,h
Second dimension 4 - 18 are 7 hand keypoint x and y coordinates: x1,y1,x2,y2,...x7,y7
"""
out_reg = self.interp_palm.get_tensor(self.out_reg_idx)[0]
"""
out_clf shape is [number of anchors]
it is the classification score if there is a hand for each anchor box
"""
out_clf = self.interp_palm.get_tensor(self.out_clf_idx)[0, :, 0]
# finding the best prediction
probabilities = self._sigm(out_clf)
detecion_mask = probabilities > 0.5
candidate_detect = out_reg[detecion_mask]
candidate_anchors = self.anchors[detecion_mask]
probabilities = probabilities[detecion_mask]
if candidate_detect.shape[0] == 0:
print("No hands found")
return None, None, None
# Pick the best bounding box with non maximum suppression
# the boxes must be moved by the corresponding anchor first
moved_candidate_detect = candidate_detect.copy()
moved_candidate_detect[:, :2] = candidate_detect[:, :2] + (
candidate_anchors[:, :2] * 256)
box_ids = non_max_suppression_fast(moved_candidate_detect[:, :4],
probabilities)
# Pick the first detected hand. Could be adapted for multi hand recognition
box_ids = box_ids[0]
# bounding box offsets, width and height
dx, dy, w, h = candidate_detect[box_ids, :4]
center_wo_offst = candidate_anchors[box_ids, :2] * 256
# 7 initial keypoints
keypoints = center_wo_offst + candidate_detect[box_ids, 4:].reshape(
-1, 2)
side = max(w, h) * self.box_enlarge
# now we need to move and rotate the detected hand for it to occupy a
# 256x256 square
# line from wrist keypoint to middle finger keypoint
# should point straight up
# TODO: replace triangle with the bbox directly
source = self._get_triangle(keypoints[0], keypoints[2], side)
source -= (keypoints[0] - keypoints[2]) * self.box_shift
debug_info = {
"detection_candidates": candidate_detect,
"anchor_candidates": candidate_anchors,
"selected_box_id": box_ids,
}
return source, keypoints, debug_info