Open Xia0ben opened 4 years ago
This may be a big issue. We directly adopt the semantic information from the raw database and map it to the NYUv2 label set. I will check this issue.
Hi @Xia0ben
Thanks for your comments. I have checked the provided case and found the semantic error.
Since the semantic annotations are labeled by the human labors, the semantic annotation may contain such errors.
Best, Jia
Hi again @bertjiazheng ,
Do you mean that the semantic.png images were 'hand-painted' by humans, were they not automatically generated from the 3D data ground truth ? Or do you mean the label associated with the ceiling lamp's 3D models was wrong ?
Do you plan on having a further look into potential mislabelings and updating the dataset ?
Wish you a pleasant day ;-).
The 2D rendering is automatically generated by the rendering process. However, the semantic label for each 3D model is labeled by the human.
So I guess there should probably not be much more labeling mistakes, then ? These flying sofas were probably just an exception (I must admit, a quite funny one) ?
I think these cases are rare in our dataset.
Hello again !
I succeeded in associating the 3D bounding boxes with semantic labels, and started digging around the dataset, and while exploring the first scene, I realized that the ceiling lamps were misidentified as sofas (you can check by simply using a color picker to compare the color of the sofas in the background with the ceiling lamps in image Structured3D/scene_00000/2D_rendering/485142/panorama/full/semantic.png). I'll try to put in place some automatic detectors and report back if I find any other incongruity.
Wish you a pleasant day.