search

lucastabelini / PolyLaneNet

Code for the paper entitled "PolyLaneNet: Lane Estimation via Deep Polynomial Regression" (ICPR 2020)
https://arxiv.org/abs/2004.10924
MIT License
295 stars 74 forks source link

Question about sharing the horizon #49

Closed voldemortX closed 3 years ago

voldemortX commented 3 years ago

@lucastabelini Hi! Great codes you are providing here! But I have a question. Is the sharing of horizon between lanes a possible reason for the low upperbound performance (since the labels don't accurately indicate horizon line)? My colleague just tried directly fitting each lane in the TuSimple test set with 2-order/3-order polynomials and got almost 100% accuracy. But in your paper the upperbound is only ~98%. BTW, the FP & FN in your TABLE 3 seem rather too low for a <98% Acc. I'm not sure where the problem is.

lucastabelini commented 3 years ago

Is the sharing of horizon between lanes a possible reason for the low upperbound performance (since the labels don't accurately indicate horizon line)?

No. There's an ablation study in the paper showing that (Table V).

My colleague just tried directly fitting each lane in the TuSimple test set with 2-order/3-order polynomials and got almost 100% accuracy. But in your paper the upperbound is only ~98%.

The script used to compute the upperbound is publicly available. You can compare your implementation to ours to find out where the discrepancy is.

BTW, the FP & FN in your TABLE 3 seem rather too low for a <98% Acc. I'm not sure where the problem is.

There's no problem. The threshold for a true positive is 85% accuracy (as defined by the benchmark).

voldemortX commented 3 years ago

Thanks for the answers! I'm aware that in actual learning, sharing the horizon does not lead to worse performance (it is indeed a natural heuristic), what I mean is the influence on the upperbound, since by comparing your upperbound code with ours (also based on numpy's fitting), that is the only possible diff we found.

lucastabelini commented 3 years ago

I see. In that case, I believe it is indeed the limiting factor.

voldemortX commented 3 years ago

Thanks a lot for confirming this! I'll close this issue for now.