Hi, I noticed that during training, how you specify the (w, h, l, z_p, theta) prior for every 3D anchor in RPN is to compute the statistics across all matching ground truths which have ≥ 0.5 IoU with its corresponding 2D anchor. But I'm wondering how are you going to generate the 3D anchors in testing time if there exists no ground truth to compute these statistics?
Hi Haorui-Ji, we use the anchors generated at training for inference at test time! Hence, after the anchors have been determined at training time, they are fixed
Hi, I noticed that during training, how you specify the (w, h, l, z_p, theta) prior for every 3D anchor in RPN is to compute the statistics across all matching ground truths which have ≥ 0.5 IoU with its corresponding 2D anchor. But I'm wondering how are you going to generate the 3D anchors in testing time if there exists no ground truth to compute these statistics?
Thanks so much~