AlekBoninski/lane-detection

Lane Detection Master Thesis Work

This repository contains my master thesis work at the Sofia University's Faculty for Mathematics and Informatics.

Train network

In this project I implemented a CNN with an architecture inspired by ERFNet.

To train the network you will need to download the TuSimple dataset found here.

Training is initiated by running the ./src/train.py file with the following args:

python train.py \
  --model erfnet \
  --resume-training \
  --epochs 100 \
  --ckpt-dir ./checkpoints \
  --ckpt-save-interval 5 \
  --cuda \
  --height 720 \
  --dataset-root /data/TUSimple \
  --pretrained-encoder ./encoder_weights.pth

--model

• Description: Model architecture to use for training
• Type: str
• Choices: encoder, encoder-low-dilation, erfnet, erfnet-low-dilation
• Default: None
• Usage:

  --model [encoder|encoder-low-dilation|erfnet|erfnet-low-dilation]

--resume-training

• Description: Flag indicating if the training should resume from a saved checkpoint.
• Default: False
• Usage:

  --resume-training

--epochs

• Description: Epochs to train for.
• Type: int
• Default: 150
• Usage:

  --epochs 150

--ckpt-dir

• Description: Directory where to save the checkpoints.
• Type: str
• Default: None
• Usage:

  --ckpt-dir ./checkpoints

--ckpt-latest-name

• Description: Name of the latest checkpoint (the model will resume training from here).
• Type: str
• Default: latest.pth.tar
• Usage:

  --ckpt-latest-name latest.pth.tar

--ckpt-best-name

• Description: Name of the checkpoint that has achieved best results.
• Type: str
• Default: best.pth.tar
• Usage:

  --ckpt-best-name best.pth.tar

--ckpt-interval-name

• Description: Template for naming interval checkpoints during training. The placeholder {} will be replaced with the epoch number.
• Type: str
• Default: checkpoint-{:04}.pth.tar
• Usage:

  --ckpt-interval-name checkpoint-{:04}.pth.tar

--ckpt-save-interval

• Description: Interval (in epochs) at which to save intermediate checkpoints.
• Type: int
• Default: 10
• Usage:

  --ckpt-save-interval 10

--loss-log-step-interval

• Description: Interval (in steps) at which to log the loss during training.
• Type: int
• Default: 50
• Usage:

  --loss-log-step-interval 50

--cuda

• Description: Enables training on GPU (CUDA). This flag is used to toggle between GPU and CPU.
• Default: True
• Usage:

  --cuda

--height

• Description: Specifies the input height for the model.
• Type: int
• Default: 512
• Usage:

  --height 512

--dataset-root

• Description: Path to the root directory of the dataset.
• Type: str
• Default: ../archive/TUSimple
• Usage:

  --dataset-root ../archive/TUSimple

--pretrained-encoder

• Description: Path to the pretrained encoder weights.
• Type: str
• Default: None
• Usage:

  --pretrained-encoder ../checkpoints/encoder/latest.pth.tar