Question about the Hungarian Algorithm Implementation

[fedyhajali]

Hi @uakfdotb,
Below I quote what is reported in the paper.

"On intermediate frames, we apply the Hungarian method on M(0,k) to match detections in Dk with tracks, updating each track with the matched detection (if any)."

I would ask if get_recur_sel() function, together with tf.gather(), represents the Hungarian algorithm implementation.

uns20/model.py

Lines 366 to 381 in 510833a

	def get_recur_sel(mat):
	if options.get('simple_sel', False):
	return tf.argmax(mat, axis=1, output_type=tf.int32)
	def f(mat):
	# take argmax along rows (over columns)
	# but only use it if it is higher value than other rows in same column
	row_argmax = numpy.argmax(mat, axis=1)
	col_argmax = numpy.argmax(mat, axis=0)
	out = row_argmax
	for i in range(out.shape[0]):
	if col_argmax[out[i]] != i:
	out[i] = mat.shape[1]-1
	return out.astype('int32')
	sel = tf.py_func(f, [mat], tf.int32, stateful=False)
	return sel

uns20/model.py

Lines 456 to 466 in 510833a

	n_prev = self.n_image[batch, 0]
	if options.get('follow_longim', False):
	sel = get_recur_sel(extra_mats_finesp[batch][prev_idx-1])
	else:
	sel = get_recur_sel(finesp_mats[batch][-1])
	rnn_sel = tf.stack([
	tf.range(n_prev, dtype=tf.int32),
	sel,
	], axis=1)
	prev_features = tf.gather(features[batch][prev_idx], sel, axis=0)
	rnn_features = tf.gather_nd(finesp_hiddens[batch][-1], rnn_sel)

[uakfdotb]

Yes, this is the part that matches detections with track prefixes during training. It is not the Hungarian algorithm but an approximation in this implementation since doing the bipartite matching is more complex and unnecessary.