scale = 700

[wangbingok1118]

MonoDet/projects/MonoRCNN/lib/dis_head.py

Line 98 in ff3da33

self.scale = 700.0

这里有什么依据吗？ 感谢感谢

[Rock-100]

1/h may be very small, so the network cannot predict it well. Multiplying 1/h with 700 can make the training more stable. In my experiments, multiplying 1/h with 400, 500, 600, or 700 all work well.