Fix potential wrong reg_target assignment on the center point collision problem as Objects as Points 6.1.1 said.
Though we cannot forbid the center point collision problem, we can guarantee that the reg_target assignment behavior is deterministic, assigning to the nearest one.
I'm not sure about the effect of this change, I have no enough machines to test it.
But I think this change is reasonable, especially for large epochs training.
By the way, I also wonder it is reasonable or not to remove this line dist2[is_peak] = 0 and believe the probability totally.
Wonderful work! CenterNet2 tells us the power of probabilistic on two-stage detection.
But the CenterNet* could be further optimized, such as ground truth auto-assignment on feature pyramid instead of configuring scale range manually.
Fix potential wrong reg_target assignment on the center point collision problem as Objects as Points 6.1.1 said. Though we cannot forbid the center point collision problem, we can guarantee that the reg_target assignment behavior is deterministic, assigning to the nearest one.
I'm not sure about the effect of this change, I have no enough machines to test it. But I think this change is reasonable, especially for large epochs training.
By the way, I also wonder it is reasonable or not to remove this line
dist2[is_peak] = 0and believe the probability totally.Wonderful work! CenterNet2 tells us the power of probabilistic on two-stage detection. But the CenterNet* could be further optimized, such as ground truth auto-assignment on feature pyramid instead of configuring scale range manually.