Hyperparameters in paper vs code

[yvanyin]

Hyperparameters in paper vs code

[YvanYin]

1. In the scaled depth space, I have tried it. It can also work. We used this configure in our later paper 'DiverseDepth: Affine-invariant Depth Prediction Using Diverse Data.' (github [https://github.com/YvanYin/DiverseDepth]) .
2. delta_diff_x/y/z is the distance difference between 2 points along x/y/z axis. delta_cos is the cosine is the cos<AB, AC>. delta_z is the threshold for the valid depth valid. z > delta_z is valid. I didn't do many ablation studies to compare different values for these hyperparameters. I just want to make the sampled points are far from each other instead of locating in a local region.

[eriksandstroem]

Thank you very much for your quick reply.

1. So this means that the variables gt_depth and pred_depth are assumed to be in the range 0-1, right? And this, in turn, means that the variables delta_z, delta_diff_x/y/z are also in this space and if I were to use this loss in the metric space e.g. let's say I assume that gt_depth and pred_depth are in the range 0-10 meters, I would want to multiply delta_z, delta_diff_x/y/z with 10, right?
2. So to confirm, according to the paper, you constrain the L2 norm between sampled points to be at least 0.6 m, but in the code you instead check that along each axis the distance is larger than delta_diff_x/y/z? Just as a sanity check: you use delta_diff_x/y/z = 0.005. If we transform this hyperparameter into the metric space 0-10m according to the reasoning in "1.", we get delta_diff_x/y/z = 0.05 m. This value does not reflect the theta= 0.6 m that is reported in the paper and 0.05 m seem like a very low number to me. Could you perhaps educate me on why you select such a low number for delta_diff_x/y/z? Also, you say delta_cos is the angle restriction of cos<AB, AC> i.e. cos<AB, AC> needs to be less than 0.867 = cos(30 degrees), but more than -0.867 = cos(150 degrees). So alpha = 150 degrees in the code, correct?

Again, thanks for your kind reply!
Cheers,
Erik

[eriksandstroem]

Hi YvanYin,
I just wanted to send you a reminder here, in case my previous message got lost. Take your time though, no rush.

Best,
Erik

[YvanYin]

Hi

1. In the training, the gt_depth is normalized to 0-1. If the gt_depth is in the 0-10m range, you have to adjust delta_z, delta_diif_x/y/z.
2. In our naive implementation, we set the distance to 0.6m. In the released code, we set the delta_diiff_x/y/z to make sampled points far away from each other. You can also set the distance between them, but these two methods are the same. We didn't do the ablation study to compare them. About the alpha, the angle between <ab, ac> is [30, 150].