`anchor_assign = False` in multi-frame with iter
FiveLu opened this issue · comments
Hi, i am so confused that why the Output-AUX branch uses anchor assign
while the Ouput branch uses proposal assign
?
with torch.no_grad():
if self.anchor_assign:
positives_mask, negatives_mask, target_positives_indices = self.assigner.match_proposals_with_targets(
anchors, target)
else:
positives_mask, negatives_mask, target_positives_indices = self.assigner.match_proposals_with_targets(
proposals[:, :5+self.anchor_len*3], target)
Thanks !
It is merely a technical trick to assign labels according to anchors or proposals. Since the output predictions from the initial iteration do not fit well with the 3d lanes, we assign their labels according to the associated anchors for training stability. For the iterative predictions, they are generated based on predictions from the previous iteration and already fit well with the 3d lanes, thus we assign their labels directly according to the predicted proposals.