search

facebookresearch / clevr-dataset-gen

A Diagnostic Dataset for Compositional Language and Elementary Visual Reasoning
Other
577 stars 204 forks source link

Inconsistency in 3d coordinates #20

Closed christopher-beckham closed 5 years ago

christopher-beckham commented 5 years ago

Hi,

I am noticing an inconsistency between the placement of the 3d objects in the scene and their corresponding 3d coordinates in the scene. For instance, to show some examples:

(1):

(it looks like the z-axis in this plot corresponds to how close/far you are from the camera, e.g. 'small-cyclinder-brown' and 'small-sphere-gray' are at the top of the z-axis, and they are closest to the camera)

(2):

(this contradicts (1) now, because the small red cylinder lies around the midpoint of the z-axis yet it's actually the furthest from the camera)

(3):

(again, contradicts the previous images, because the gold sphere is closest to the camera but lies about the mid-point of the z-axis)

I have not generated the dataset myself from source, but if there is a discrepancy between the code and the dataset that is available online that may explain it. Or, if the 3d coordinates are not relative to the camera (e.g. the camera was randomly displaced prior to rendering). I tried looking at the code but I didn't find anything that appeared to be odd to me, but I am really not sure at this stage.

Thanks!

christopher-beckham commented 5 years ago

nvm closing this, the z values were actually corresponding to the vertical axis and things were much clearer if I made all of those values zero. In other words, there are no issues with the coords.