A semantic segmentation baseline using @comma.ai's comma10k dataset.
Using U-Net with efficientnet encoder, this baseline reaches 0.044 validation loss.
Here is an example (randomly from the validation set, no cherry picking)
This baseline uses two stages (i) 437x582 (ii) 874x1164 (full resolution)
python3 train_lit_model.py --backbone efficientnet-b4 --version first-stage --gpus 2 --batch-size 28 --epochs 100 --height 437 --width 582
python3 train_lit_model.py --backbone efficientnet-b4 --version second-stage --gpus 2 --batch-size 7 --learning-rate 5e-5 --epochs 30 --height 874 --width 1164 --augmentation-level hard --seed-from-checkpoint .../efficientnet-b4/first-stage/checkpoints/last.ckpt
- Update to pytorch lightning 1.0
- Try more image augmentations
- Pretrain on a larger driving dataset
- Try over sampling images with small or far objects
Python 3.5+, pytorch 1.6+ and dependencies listed in requirements.txt.