why not the bias depth is a negative number?

[lfydegithub]

bias_depth = 1. / (depth_net_out[:,0:1].sigmoid() + 1e-6) - 1. ∈(0, +inf)
the bias_depth need be negative number if the geo_depth was bigger than real_depth.

[SuperMHP]

it is right. Because we use box 2d height as h2d, so the projection depth will be smaller than targets. So the bias should be also larger than zero.

…

在 2022年1月13日，下午4:06，pony ***@***.***> 写道：  bias_depth = 1. / (depth_net_out[:,0:1].sigmoid() + 1e-6) - 1. ∈(0, +inf) the bias_depth need be negative number if the geo_depth was bigger than real_depth. — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you are subscribed to this thread.

[llwc]

Hello, I have a question about bias_depth = 1. / (depth_net_out[:,0:1].sigmoid() + 1e-6) - 1? what is the difference with using bias_depth = exp(-depth_net_out[:,0:1])? Another question, this bias_depth cannot fixed the depth error if the bbox 2d height is smaller than its expected. Thanks!

[SuperMHP]

Hello, I have a question about bias_depth = 1. / (depth_net_out[:,0:1].sigmoid() + 1e-6) - 1? what is the difference with using bias_depth = exp(-depth_net_out[:,0:1])? Another question, this bias_depth cannot fixed the depth error if the bbox 2d height is smaller than its expected. Thanks!

1). It is equal. This is provided by our baseline code implementation.
2). The reason of bias>0 can be seen at the answer before your question.