Open ddannylee opened 3 years ago
Hi @ddannylee,
The starting reason for this choice is because of the orientation regression. If we allow yaw value to be from [0, 2pi), and use a loss such as ||yaw-yaw_gt||, then yaw=0.1 and yaw_gt=2pi have a large loss. We should mod the angle in the loss. To avoid this behavior, we can just break the angle into different bins.
Also, it's believed regression on multimodal is hard and therefore breaking the regression into small ranges is helpfuling. e.g. anchors are often used in 2D detection.
I didn't dig much into it. Hope it helps, thank you.
Hi!
i have a question about your implementation details.
you had set two different class that one is front view, another is side view!
Can you teach me why you set a different class?
Thank you