How to convert pose to utm about patch-netvlad

[Tobias-Fischer]

Hi,

utmQ and utmDb are simply N*2 numpy arrays where each row contains the x & y coordinates of a query / reference place. The posDistThr scalar is the ground truth tolerance (typically in meters). See

Patch-NetVLAD/patchnetvlad/tools/datasets.py

Lines 64 to 65 in 4edee2f

	if ground_truth_path is not None:
	self.utmQ, self.utmDb, self.posDistThr = self.parse_gt_file(ground_truth_path)

Patch-NetVLAD/patchnetvlad/tools/datasets.py

Lines 93 to 102 in 4edee2f

	def get_positives(self):
	# positives for evaluation are those within trivial threshold range
	# fit NN to find them, search by radius
	if self.positives is None:
	knn = NearestNeighbors(n_jobs=-1)
	knn.fit(self.utmDb)
	self.distances, self.positives = knn.radius_neighbors(self.utmQ, radius=self.posDistThr)
	return self.positives

Patch-NetVLAD/patchnetvlad/tools/datasets.py

Lines 120 to 124 in 4edee2f

	@staticmethod
	def parse_gt_file(gtfile):
	print('Parsing ground truth data file...')
	gtdata = np.load(gtfile)
	return gtdata['utmQ'], gtdata['utmDb'], gtdata['posDistThr']

Patch-NetVLAD/feature_match.py

Lines 153 to 160 in 259874b

	if opt.ground_truth_path is not None:
	print('Calculating recalls using ground truth.')
	gt = eval_set.get_positives()
	global_recalls = compute_recall(gt, predictions, eval_set.numQ, n_values, 'NetVLAD')
	local_recalls = compute_recall(gt, reranked_predictions, eval_set.numQ, n_values, 'PatchNetVLAD')
	write_recalls_output(opt, global_recalls, local_recalls, n_values)

for relevant code snippets.

[DonMuv]

But my ground truth file, the poses of the images come in the format of a 7 feature vector (t_x, t_y, t_z, q_x, q_y, q_z, q_w), with t corresponding to the translation vector and q to the rotation expressed in form of a quaternion.

such as
[ 4.45019497e+00 8.69529027e+00 4.75170998e-01 6.67300280e-01 -2.33674360e-01 2.27086490e-01 6.69730062e-01]
[... ... ]
...

pose matrix of shape = (n, 7)

[Tobias-Fischer]

Hi,
Simply modify the get_positives() method above to define your desired ground truth tolerance. In our case, it's simply the Euclidean distance in the plane; if you want to use this simply disregard everything except for t_x and t_y (sth like utmQ = utmQ[:, :2] and the same for utmDb)

Alternatively, you could adapt get_positives to separately have some tolerance for position (based on t) and orientation (based on q). You should be able to use some code from e.g. kapture to do this. utmQ and utmDb will be a bit confusing naming of the variables, as they will contain a matrix of size N*7, but otherwise everything else should work the same.

[Tobias-Fischer]

Feel free to re-open if you have more questions.