search

MarlonCajamarca / Keras-LSTM-Trajectory-Prediction

A Keras multi-input multi-output LSTM-based RNN for object trajectory forecasting
114 stars 15 forks source link

Dataset #1

Closed JiajianLu closed 5 years ago

JiajianLu commented 5 years ago

Hi, can you also upload the link to download your training and testing data? Thank you!

MarlonCajamarca commented 5 years ago

Hi @JiajianLu Of course! You can download a raw dataset here. Once you download the raw_dataset, you can create a ready-to-use training/test train_test.hdf5 dataset by using opt_train_test_dataset_transformer.py or train_test_dataset_transformer.py scripts, like this:

python3 train_test_dataset_transformer.py /Path/to/input_raw_dataset.hdf5 /Path/to/output/dataset/folder/<output_train_dataset.hdf5> config_transformer.json

Then, you could use the different training and inference scripts (train.py, train_DrBN.py) to play around with different model configurations and hyperparameters. Please check the ReadMe.md file for detailed documentation about the different tools and corresponding Command-Line-Interface examples executions. Have fun!

JiajianLu commented 5 years ago

Thank you @MarlonCajamarca! Does the dataset include type of objects (e.g. pedestrian, car, bicycle) and do you consider it as one of the inputs to the network?

MarlonCajamarca commented 5 years ago

Hi @JiajianLu

Sorry for the late response. This is the workflow now: We use the raw_dataset_creator tool for generating a raw_dataset.h5 file containing all the sample sequences, it does not matter the trajectory_length or the type of vehicle performing that trajectory. However, in this raw_dataset.h5 each of the appended dataset/trajectory sample has its own attribute which captures the type of vehicle performing that sample trajectory (see this link for more info) for further filtering processes.

Then, we used one of the train_test_transformator tools for creating the final train/test datasets where we filter each of the input trajectories and keep only the trajectories associated with a specific or group of specific vehicle classes using the classes_2_extract hyperparameter available in the config_transform.json file.

For example, classes that have been used for creating the datasets are: classes_2_extract = ["LIGHTWEIGHT_AUTOMOTIVE_VEHICLE", "BUS", "VAN"]

We can also create a trajectory predictor for PEDESTRIANS, MOTORCYCLES, and so on. However, the actual datasets do not comprise these additional classes yet. I will be adding various datasets composed by other classes apart from these first baseline categories, so keep in touch for the incoming updates.

Regards, Marlon!