Inconsistency with inference angles

[heechunlee]

Hi! I was using this project to train a custom dataset but realized that there were inconsistencies in the angle given by cos theta and sin theta values in the outputted labels. So in case I was making an error formatting my own labels, I ran rotate_detect.py on the UCAS-AOD images. Below is the outputted label of 1103.png: Recalculating the arccos and arcsin of the cos theta and sin theta values still seems to give different values of the angle. Any idea why this is? Thanks

[XinzeLee]

I am not sure I get your point. You may refer to the following points for reference.

1. arcsin and arccos give the outputs in radians, not in degree;
2. To represent the angle (which lies within -90 to 90), two values should be used, cos_theta and sin_theta. To convert it back with one function, you may consider arctan (which is arcsin / arccos).

[StarBlue98]

Thank you for your excellent work! I've learnt a lot from it! I want to use this model to train my own dataset, but I'm not sure what the right format of the labels are. You use UCAS-AOD dataset, and the labels, for example the first row of 22.txt are as follows: 0 0.255303 0.319417 0.093962 0.099477 0.493946 -0.869493 the first five num means class, x, y, w, h, respectively. I get confused about what the last two num represent? And when I use this model to train my own data, should I modify my labels like (x, y, w, h, re, im)? or the format like (cx, cy, w,h, angle) or (x1,y1,x2,y2,x3,y3,x4,y4) would be OK?

[XinzeLee]

To clarify, for the example label "0 0.255303 0.319417 0.093962 0.099477 0.493946 -0.869493", it represents cx, cy, width, height, cosine_theta, sine_theta respectively, so the last two numbers represent cosine_theta and sine_theta.

For your own dataset, the label formats should be consistent with the examples provided, i.e. cx, cy, width, height, cosine_theta, sine_theta.

[heechunlee]

Sorry I haven't gotten back on this. What I meant to say was, since cos_theta and sin_theta are computed from the same theta value, then wouldn't applying arccos and arcsin to the cos_theta and sin_theta values, respectively, return the same angle? So in the case of the fourth line of the 1103.png label ("0 0.108594 ... 0.989707 0.073235"), arccos(0.989707) = 0.1436 radians while arcsin(0.073235) = 0.0733 rad. This is a difference of a little over 4 degrees, so I was wondering where this inconsistency was originating from? Please let me know if there are problems with my understanding, thanks!

[StarBlue98]

@heechunlee

few days ago, I have similar confusion. I hope the following code would help you.

theta = math.acos(cos)*180/np.pi

求theta：math.acos(cos)* 180/np.pi

求sin：math.sin(theta* np.pi/180)

By the way, I want to use my own label like (445.8136 222.3948 120.8183 123.9594), which means cx, cy, w, h in pixel, so I wonder how to normalize it?