About roll angle rectify after image flipping

[sunnyykk]

Hello,

I noticed that in line 159, you employed random flipping to augment the query and map images, and later, in line 164, the query image are rectified by setting the roll and pitch angles to zero. However, I'm wondering if a negative sign should be applied to the roll angle during the rectification process after the images have been flipped.

I'm quite new to this area, thank you for your attention to this detail!

[sunnyykk]

To further elaborate on the issue mentioned above, some visualizations are provided:

• The first image below is a training image after random_flip function, which includes mirrored Chinese characters;
• The following image is after gravity correction using the process_image function; note that the lateral roll angle has become larger;
• The third image is the result after I took the negative value of the roll angle and reprocess it;

Is it possible that there is a bug in the gravity correction process during training that might potential affect half of the augmented training samples?

[sarlinpe]

You are right, thank you investigating and reporting this! I likely missed it on visual inspection because most images of our dataset have zero roll. I still think that fixing it will improve the model. Could you submit a PR for it? I'll retrain the model once I find the time.

[sunnyykk]

thanks! I will submit it~