Closed JuanSeBestia closed 5 years ago
Pose is another task altogether, different to object detection, where a different model than Faster R-CNN is required. This task is normally called pose estimation, and Luminoth does not yet support it.
As of the other attributes, they have always been ignored in the official Pascal VOC challenge:
The difficult field being set to 1 indicates that the object has been annotated as ``difficult'', for example an object which is clearly visible but difficult to recognize without substantial use of context. Currently the evaluation ignores such objects, contributing nothing to the precision/recall curve.
More than help during training, these likely add noise that confuses the algorithms that currently exist.
Thanks for answering
I know that in other formats it is not supported, but in VOC yes the attributes of truncated, difficult, pose, I understand that the pose may not be very easy to parameterize, but the other two do not.
Do you think that these parameters can help a better train?