search

cfzd / Ultra-Fast-Lane-Detection

Ultra Fast Structure-aware Deep Lane Detection (ECCV 2020)
MIT License
1.82k stars 493 forks source link

How to build my own data set and modify network settings then train it #112

Closed CNSTJiatongLee closed 4 years ago

CNSTJiatongLee commented 4 years ago

I want to label my own data, but I don’t know the supported label format requirements. At the same time, I want to know how to modify the settings to fit my own data set. My expectation is that the model can extract the center trajectory of each lane. img label label_viz

cfzd commented 4 years ago

@CNSTJiatongLee Basically, our method supports the CULane dataset, so if your dataset fits the format of CULane, you can directly use our code to train on your data.

The requirements of CULane:

CNSTJiatongLee commented 4 years ago

@CNSTJiatongLee Basically, our method supports the CULane dataset, so if your dataset fits the format of CULane, you can directly use our code to train on your data.

The requirements of CULane:

  • lanes should be continuous line segmentation map
  • each lane is a separate class
  • CULane defines one background class and 4 valid classes (i.e. four kinds of lanes), which are leftmost, left, right, rightmost
  • The total range of label is [0-5], 0 represents the background, [1-4] represents the four kinds of lanes

Does it support the Tusimple data set format? I am using the labelme tool for point labeling

CNSTJiatongLee commented 4 years ago

@CNSTJiatongLee Basically, our method supports the CULane dataset, so if your dataset fits the format of CULane, you can directly use our code to train on your data.

The requirements of CULane:

  • lanes should be continuous line segmentation map
  • each lane is a separate class
  • CULane defines one background class and 4 valid classes (i.e. four kinds of lanes), which are leftmost, left, right, rightmost
  • The total range of label is [0-5], 0 represents the background, [1-4] represents the four kinds of lanes

Does it support the Tusimple data set format? I am using the labelme tool for point labeling

cfzd commented 4 years ago

@CNSTJiatongLee Yes, we also support Tusimple format. We provide a tool scripts/convert_tusimple.py that could convert the Tusimple format to the trainable CULane format.