search

XingangPan / SCNN

Spatial CNN for traffic lane detection (AAAI2018)
MIT License
810 stars 242 forks source link

The lane detection model is the multi-label and multi-task problem? #34

Open FrancisGee opened 6 years ago

FrancisGee commented 6 years ago

The lane detection model is the multi-label and multi-task problem? And I really want to know what the probmaps means

XingangPan commented 6 years ago

@FrancisGee The lane detection model has two branches. The problem could be viewed as a multi-task problem. The last convolution layer of the model would output feature maps of 5 channels. When the 5 channels go through a softmax layer, they form the probmaps. Because for each pixel location, the values of the 5 channels indicate the probability that the pixel belongs to the 5 categories (background, lane marking 1-4).

OneMondy commented 6 years ago

@XingangPan Hello, I would like to ask which layer of the SCNN network is specifically added in the Figure 5 model. What does the interp*8 in Figure 5 mean? The paper points out that the model includes 20 layers, 16 layers of convolution and 4 layers of SCNN layers. Is this correct? Looking forward to your answers, thanks

XingangPan commented 6 years ago

@OneMondy

  1. SCNN is added between fc7 and fc8 in Fig.5.
  2. Interpx8 means enlarge the featuremap by 8 times (along height and width dimensions) with bilinear interpolation.
  3. Yes, it's correct.