Several tricks to boost the performance of SCNN-Tensorflow in CULane

[cardwing]

There are several tricks to boost the performance of SCNN-Tensorflow:

• Decrease the coefficient of the lane existence prediction loss
• Decrease the coefficient of the background pixels in the cross-entropy loss
• First pre-train the VGG-16 backbone on CULane. Then add the message passing module to VGG-16 and train the whole model jointly
• Introduce auxiliary tasks like drivable area detection and lane point regression
• Use a better optimizer, like SGD + Nesterov Momentum
• Balance the proportion of different driving categories like normal, shadow and curve, in the training process
• Hard example mining
• Add weight decay