While reading through the dataset preparation I stumbled upon the piece of code above, which is (I assume) used to perform depth augmentation as outlined in your publication.
The intention is clear: perturbing both the point cloud and the box by a function of the distance by sampling from a unit gaussian and multiplying the result by 0.05*dist.
The problem lies in the clipping operation that follows: since the clipping range is set to [0.8*dist, 1.2*dist], the result is that the value is always equal to 0.8*dist, since the sampled value is always much smaller.
Is this intended or is this a bug?
Another question: why consider only the xy distance instead of the full distance? This value tends to be very small.
While reading through the dataset preparation I stumbled upon the piece of code above, which is (I assume) used to perform depth augmentation as outlined in your publication.
The intention is clear: perturbing both the point cloud and the box by a function of the distance by sampling from a unit gaussian and multiplying the result by 0.05*dist.
The problem lies in the clipping operation that follows: since the clipping range is set to
[0.8*dist, 1.2*dist]
, the result is that the value is always equal to0.8*dist
, since the sampled value is always much smaller.Is this intended or is this a bug?
Another question: why consider only the xy distance instead of the full distance? This value tends to be very small.
Thank you.