How TAO training deals with occlusions

• Hardware (T4)
• Network Type (Detectnet_v2/Yolo_v4/)

Hi, I would like to know how detectnet_v2 deals with occluded bboxes. I know that KITTI format has a parameteres to set it, but in the training spec file there is no information about that, in the documentation neither. Maybe it is treated as a normal bbox or inside the tensorflow has a special treatment idk. I’m very curious on that.

Same case for yolo_v4, it has a parameter include_difficult_in_training=True, but… that’s it?

Thanks in advance.

Yes. For yolov4, please refer to YOLOv4 — TAO Toolkit 3.22.02 documentation

include_difficult_in_training : A flag specifying whether to include difficult boxes in training. If set to false , difficult boxes will be ignored. Difficult boxes are those with non-zero occlusion levels in KITTI labels.

For detectnet_v2, it is not needed to set in the spec file. The label file already includes it. Data Annotation Format — TAO Toolkit 3.22.02 documentation

For detectnet_v2 if I fill the parameter occlussion=2, what does it mean? what TAO train will do? treat like occlusion or like another bbox without occlusion?

For detectnet_v2, it will do nothing for this kind of label. So, the occlusion setting takes no effect in detectnet_v2 training.

I got it. And for yolo_v4, do you know what kind of process is done for occlusions at the moment of training?

For yolov4, if set occlusion to 2 in label, these difficult boxes will be included in training. If set to include_difficult_in_training: false , difficult boxes will be ignored.

yeah I know, but how those oclussions are treated at training?

Sorry for late reply. As mentioned above, if set to include_difficult_in_training: false , difficult bboxes will be ignored. That label will be excluded. If set to include_difficult_in_training: True , difficult bboxes will be considered. That means the label will be used.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.