Data augmentation possibly falsify some answers?

[jeasinema]

Hi,

I found the following data augmentation in the dataset pipeline

ScanQA/lib/dataset.py

Lines 291 to 299 in 422dde5

	if np.random.random() > 0.5:
	# Flipping along the YZ plane
	point_cloud[:,0] = -1 * point_cloud[:,0]
	target_bboxes[:,0] = -1 * target_bboxes[:,0]
	if np.random.random() > 0.5:
	# Flipping along the XZ plane
	point_cloud[:,1] = -1 * point_cloud[:,1]
	target_bboxes[:,1] = -1 * target_bboxes[:,1]

I believe this is from the original VoteNet code base, which "literally" mirrors the point cloud. That makes sense for 3D object detection as long as it also mirrors the bbox coordiates. However, such augmentation could completely change the relative relations among objects (A is to the left of B -> A is to the right of B, as the objects are also "mirrored"), thereby falsify some answers. My questions are:

-Did you compare results of training w/ or w/o such data augmentation?
-What do you think of removing this from the code base as it might be a bit misleading, even if it can be disabled via --no_augmentation?

[miyatai2]

Hi,

For the first question, I experimented with and without data augmentation. As far as I remember, there was no significant difference in overall performance. But, if you look at each question in detail, there may be differences.

For the second question, data augmentation may be useful for training the backbone, so I leave it as is. I would leave it to the user to decide whether to use the data augmentation or not since it can be easily switched in the options, as you have pointed out.