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

The performance seems to be bad with my datasets. #11

Closed feifeiwei closed 4 years ago

feifeiwei commented 4 years ago

hello, I have trained polyLanenet with my data, but the results performance are bad. the epochs are 250, the poly loss is around 0.1

lucastabelini commented 4 years ago

How many images does you dataset have? What settings did you use? There are many things that will affect the model's performance. Maybe if you provide more information I may be able to help you.

feifeiwei commented 4 years ago

There are about 35000 training images. My setting are as follows: 

backbone: mobilenetv2
conf_weight: 1
  lower_weight: 1
  upper_weight: 1
  cls_weight: 0
  poly_weight: 300
  batchsize: 48
optimizer:
  name: Adam
  parameters:
    lr: 3.0e-4
img_size: h, w = [256, 512]
lucastabelini commented 4 years ago

You're using a backbone that I did not experiment with, so I can't say whether it should work or not. Which learning rate scheduler are you using? The one I used has a minimum at 385 * (n + 1) epochs. If you're using the same, in the epoch 250 it will not be at a minimum, and that can affect the model's performance, since the learning rate will still be quite high. Do you have any performance metrics on the test set? I can't really tell how bad it is based on 3 images.

Apart from that, I can't really help much without access to the code and data you're using.

feifeiwei commented 4 years ago

Thank your for your help, I would try another backbone, hope it will help.