This project accompanies the paper Periphery-Fovea Multi-Resolution Driving Model Guided by Human Attention (https://arxiv.org/abs/1903.09950)

The code was written with Tensorflow 1.5, a customized version of Keras 2.1.5 and some other common packages. A Docker image blindgrandpa/periphery_fovea_driving_model was prepared for running the code. The Dockerfile of that Docker image is at ./docker_images/periphery_fovea_driving_model/ in this repo. The Dockerfile lists all the dependencies. In order to use this Docker image to run our code, you need to have nvidia-docker installed.

1. Please visit this GitHub repo and follow the instructions there to prepare the data in BDD-V for our model.

2. Download the pre-trained attention prediction model and driving model. Download this zip file and unzip it to ./. After unzip, you should see ./attention_prediction_model_checkpoints and ./logs/pre-trained_model.

3. Download the pre-trained weights of Alexnet. Downlaod bvlc_alexnet.npy and put it at ./.

4. Start training by running the following command. Replace a_name_for_this_experiment with another folder name you like.

python3 train.py \
--data_dir=data \
--model_dir=logs/a_name_for_this_experiment \
--use_foveal=True \
--attention_model_dir=attention_prediction_model_checkpoints/pre-trained \
--sample_fovea=True \
--attention_logit_factor=1 \
--batch_size=3 \
--n_steps=20 \
--validation_batch_size=3 \
--validation_n_steps=20 \
--train_epochs=220 \
--epochs_before_validation=1 \
--feature_map_channels=256 \
--quick_summary_period=120 \
--slow_summary_period=1200 \
--augment_data=False \
--small_camera_size 72 128 \
--camera_size 720 1280 \
--gazemap_size 9 16 \
--n_future_steps=10

5. To track the training in Tensorboard, run the following command.

tensorboard --logdir=logs/a_name_for_this_experiment

1. The original videos of BDD-V dataset are about one-minute long. Your machine may not be able to fit one whole video in. Run the following command to divide the testing videos into segments of a length of your choice. Note that the unit of --max_length is the number of frames. In this example command, --max_length is set to 310. This is because the frame rate is 10Hz, the model predicts for one second in the future and we would like to test the model over 30-second-long sequences.

python3 divide_tfrecords.py \
--data_dir=data \
--visible_gpus=0 \
--max_length=310

2. Run the following command to evaluate a model. Replace a_name_for_this_experiment with the folder name that you set in training, or replace it with pre-trianed_model to test our pre-trained model. If you did not divide the test videos into short segments, set --multiple_tfrecords=False. The program will create two Pandas data frames, outputs.feather and videos.feather in the folder ./logs/a_name_for_this_experiment/prediction_iter_*. The data frame outputs.feather records the predicted speed and prediction error at each frame and videos.feather records the videos used and their keys.

python3 predict_and_evaluate.py \
--data_dir=data \
--model_dir=logs/a_name_for_this_experiment \
--use_foveal=True \
--attention_model_dir=attention_prediction_model_checkpoints/pre-trained \
--sample_fovea=True \
--attention_logit_factor=1 \
--batch_size=1 \
--feature_map_channels=256 \
--small_camera_size 72 128 \
--camera_size 720 1280 \
--n_future_steps=10 \
--multiple_tfrecords=True