What is the pose tracking method used in this simple-HRNet

[jellyfish1456]

Hi~ Thank you for sharing this awesome job

I wanna ask that what is the tracking method you add in simple-HRNet?
Because the original method and paper seems to be no tracking method in it?

Thank you~

[stefanopini]

Hi!

It is a trivial tracking by detection algorithm, defined here:

simple-HRNet/misc/utils.py

Lines 382 to 425 in f4a8174

	def find_person_id_associations(boxes, pts, prev_boxes, prev_pts, prev_person_ids, next_person_id=0,
	pose_alpha=0.5, similarity_threshold=0.5, smoothing_alpha=0.):
	"""
	Find associations between previous and current skeletons and apply temporal smoothing.
	It requires previous and current bounding boxes, skeletons, and previous person_ids.
	Args:
	boxes (:class:`np.ndarray`): current person bounding boxes
	pts (:class:`np.ndarray`): current human joints
	prev_boxes (:class:`np.ndarray`): previous person bounding boxes
	prev_pts (:class:`np.ndarray`): previous human joints
	prev_person_ids (:class:`np.ndarray`): previous person ids
	next_person_id (int): the id that will be assigned to the next novel detected person
	Default: 0
	pose_alpha (float): parameter to weight between bounding box similarity and pose (oks) similarity.
	pose_alpha * pose_similarity + (1 - pose_alpha) * bbox_similarity
	Default: 0.5
	similarity_threshold (float): lower similarity threshold to have a correct match between previous and
	current detections.
	Default: 0.5
	smoothing_alpha (float): linear temporal smoothing filter. Set 0 to disable, 1 to keep the previous detection.
	Default: 0.1
	Returns:
	(:class:`np.ndarray`, :class:`np.ndarray`, :class:`np.ndarray`):
	A list with (boxes, pts, person_ids) where boxes and pts are temporally smoothed.
	"""
	bbox_similarity_matrix, pose_similarity_matrix = compute_similarity_matrices(boxes, prev_boxes, pts, prev_pts)
	similarity_matrix = pose_similarity_matrix * pose_alpha + bbox_similarity_matrix * (1 - pose_alpha)
	m = munkres.Munkres()
	assignments = np.asarray(m.compute((1 - similarity_matrix).tolist())) # Munkres require a cost => 1 - similarity
	person_ids = np.ones(len(pts), dtype=np.int32) * -1
	for assignment in assignments:
	if similarity_matrix[assignment[0], assignment[1]] > similarity_threshold:
	person_ids[assignment[0]] = prev_person_ids[assignment[1]]
	if smoothing_alpha:
	boxes[assignment[0]] = (1 - smoothing_alpha) * boxes[assignment[0]] + smoothing_alpha * prev_boxes[assignment[1]]
	pts[assignment[0]] = (1 - smoothing_alpha) * pts[assignment[0]] + smoothing_alpha * prev_pts[assignment[1]]
	person_ids[person_ids == -1] = np.arange(next_person_id, next_person_id + np.sum(person_ids == -1))
	return boxes, pts, person_ids

In short, it builds a similarity matrix between the previous and the current detections (based on bbox and pose similarity) then it associates previous and current detections with the Munkres algorithm to maximize the similarity of the associations.