search

sekilab / VehicleOrientationDataset

The vehicle orientation dataset is a large-scale dataset containing more than one million annotations for vehicle detection with simultaneous orientation classification using a standard object detection network.
MIT License
51 stars 6 forks source link

Egocentric vs. Allocentric #2

Closed skmhrk1209 closed 11 months ago

skmhrk1209 commented 11 months ago

Are the annotated orientation labels egocentric (orientation w.r.t. camera) ? or allocentric (orientation w.r.t. object)? You can refer to the 3D-RCNN paper for the details. It is very important when an orientation classifier is trained on cropped images. Thank you.

ashutoshIITK commented 11 months ago

@skmhrk1209

Thank you for your question. All the annotations are labeled from the camera's viewpoint, so the orientation labels are egocentric. Thank you!

skmhrk1209 commented 11 months ago

Thank you for your reply. Here I show an extreme example where all the vehicles have the same egocentric orientation. So, in this case, all the vehicles are annotated with the same label "front", even though their appearances are very different. Is my understanding right?

orientation drawio

ashutoshIITK commented 11 months ago

@skmhrk1209 Thank you for the clarification and sorry for my misunderstanding. According to the image you have shown, vehicles in the image will be labeled front (1 vehicle) and side (2 vehicles) orientations.

So basically, annotations are based on appearance and it should be allocentric. Sorry for my confusion earlier. Thank you!

skmhrk1209 commented 11 months ago

OK. I understand. Thank you very much!