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
301 stars 75 forks source link

Do I need to label lanes in a fixed order? #28

Closed PJHardy closed 3 years ago

PJHardy commented 4 years ago

Thanks for your sharing.

Tusimple labeled 4 lanes for one image, and they have fixed order like left-left, ego-left, ego-right, right-right from 0 to 3. Does the output of PolyLaneNet(L1 ~ L5) have the save order, or it just predict lane randomly?

If I want training my own dataset, do I need to label lanes in a fixed order?

lucastabelini commented 4 years ago

No, the code will sort the lanes for you (here).

PJHardy commented 4 years ago

So the targets were sorted in this order, will the predicts of net obey this order?

When calculate loss you simply use MCE or BCE without any matching for targets and predicts, unlike the common object detection method, they usually need to find the connection between anchor and ground truth.

If 4 lanes are labeled in a certain image and the predicts have 5 lanes, it will just use the first 4 lanes of prediction to calculate loss(from here). why does it work?

Another question, If I using tusimple dataset for training, can I just assume the predicts (L1, L2, L3, L4) are exactly (left-left, ego-left, ego-right, right-right)?

lucastabelini commented 4 years ago

So the targets were sorted in this order, will the predicts of net obey this order?

Yes, probably.

When calculate loss you simply use MCE or BCE without any matching for targets and predicts, unlike the common object detection method, they usually need to find the connection between anchor and ground truth.

If 4 lanes are labeled in a certain image and the predicts have 5 lanes, it will just use the first 4 lanes of prediction to calculate loss(from here). why does it work?

The model learns to predict all lanes in a sorted manner.

Another question, If I using tusimple dataset for training, can I just assume the predicts (L1, L2, L3, L4) are exactly (left-left, ego-left, ego-right, right-right)?

Not sure. You would have to look at the dataset's original description.