This post showcases the training result of a combination of SSD and FC8 network. Modified SSD which classifies 41 classes and FC8 network that outputs 3 types of road.

  • Training specs
    • Dataset: bdd100k Link
    • Hardware: NVIDIA Telsa P4
    • Epochs: 10
    • Time to train: 3 days
    • Training set: 70k data
    • Effort: 2 weeks of casual coding
    • Reward: I guess personal satisfaction, none monetary though
  • Samples from final training: png png png png
  • Final thoughts

Something I failed to solve to fit a polynomial line on the segmentation map. I actually just realized I should predict polynomial points on the label set instead of doing the full segmentation prediction. So ideally the network should just spits out polynomial points for the contour of the segmentation map.

Finally I didn’t put these code on my github. Perhaps it will stay on my hard drive forever. Most of the time was spend on writing miscellanous stuffs such as parsing the data and writing code for post-processing. I felt that AI needs proper engineerning more than just designing the network itself.