why using anchors but propoals when matchs proposals and gt?

[zimurui]

I an confusing when matching proposals and gt, why laneatt using anchors match but not proposals that model inference.

[lucastabelini]

Because it results in more stable training. If you use the proposals to match with the labels, the target for the same image may change as the training progresses, which can make the convergence take longer.

[nhduong1203]

When I train, sometime the regression loss is too large (2000 or 3000) while the classification loss is about 0.5 (4 class). Then I check the code and have the same question. So when the model matching, it don't care about the proposal, so sometime, it's match the anchor with the nearest lane, but the offset corresponding to the anchor is too large or something that make the proposal is much different from the lane, it cause my regression loss is large, right?
Please give me your opinion, thanks you <3.
Sorry for my English not good

[ztianlin]

If you train from scratch, proposals are randomly distributed at the beginning, which means you may get the wrong match when using proposals to match targets, making it hard for convergence. Instead, if you fine-tune from a well-predicted model, I don't see problems in using proposals.