Drivable Area Segmentation and Parameterization

PyTorch implementation of our framework for Drivable Area Segmentation and Parameterization using the ENet Architecture applied to the *Berkeley Deep Drive Dataset containing 100K drivable area maps.

Usage

Run [main.py], the script used to run the training/testing of the ENet model

python main.py [-h] [--num-epochs NUM_EPOCHS] [--learning-rate LEARNING_RATE]
               [--lr-decay LR_DECAY] [--lr-decay-epochs LR_DECAY_EPOCHS]
               [--weight-decay WEIGHT_DECAY] [--epochs EPOCHS]
               [--run-cuda CUDA] [--batch-size BATCH_SIZE]
               [--data-dir DATA_DIR] [--data-list DATA_LIST]
               [--input-size INPUT_SIZE] [--gpu-select GPU_SELECT]
               [--run-name RUN_NAME] [--mode MODE] [--load LOAD]

For help on the optional arguments run: python main.py -h

Take a look at [src/arguments.py] to check for the default arguments

Training Arguments (Example):

python main.py --gpu-select 0 --batch-size 10 --run-name train_example

(OR) Equivalently,

python main.py -g 0 -b 10 -rn train_example

Testing Arguments (Example):

Note: This example uses a model checkpoint stored in [saved_models/run4/]. If the model is elsewhere, provide the path to the model instead.

python main.py --gpu-select 0 --batch-size 5 --run-name test_example --mode test --load saved_models/run4/checkpoint_20.h5

(OR) Equivalently,

python main.py -g 0 -b 5 -rn test_example -m test -l saved_models/run4/checkpoint_20.h5

Code Organization

Directories

• [dataset] contains the dataloader codes, the dataset (not on the repository, download it yourself with the instructions given below) and list of labels that are used by the dataloader

• [src] contains the training and testing codes, the metrics that are used for evaluation in [src/metrics], the arguments and the helper functions used by the other codes.

• [models] contains the ENet model architecture code definitions

• [saved_models] contains the saved model checkpoints of our imple- mentation inside a folder corresponding to the run_name provided at run time.

Codes

• [src/arguments.py] Contains all the parsable command-line options and their defaults.
• [dataset/bdd_dataset.py] Contains the DataLoader classes for each of train,valid and test datasets
• [src/train.py] Defines the TrainNetwork class used to train the model
• [src/test.py] Defines the TestNetwork class used to test the train model