trajdata: A Unified Interface to Multiple Human Trajectory Datasets

Announcements

Sept 2023: Our paper about trajdata has been accepted to the NeurIPS 2023 Datasets and Benchmarks Track!

Installation

The easiest way to install trajdata is through PyPI with

pip install trajdata

In case you would also like to use datasets such as nuScenes, Lyft Level 5, or Waymo Open Motion Dataset (which require their own devkits to access raw data or additional package dependencies), the following will also install the respective devkits and/or package dependencies.

# For nuScenes
pip install "trajdata[nusc]"

# For Lyft
pip install "trajdata[lyft]"

# For Waymo
pip install "trajdata[waymo]"

# For INTERACTION
pip install "trajdata[interaction]"

# All
pip install "trajdata[nusc,lyft,waymo,interaction]"

Then, download the raw datasets (nuScenes, Lyft Level 5, ETH/UCY, etc.) in case you do not already have them. For more information about how to structure dataset folders/files, please see DATASETS.md.

Package Developer Installation

First, in whichever environment you would like to use (conda, venv, ...), make sure to install all required dependencies with

pip install -r requirements.txt

Then, install trajdata itself in editable mode with

pip install -e .

Data Preprocessing [Optional]

The dataloader operates via a two-stage process, visualized below. While optional, we recommend first preprocessing data into a canonical format. Take a look at the examples/preprocess_data.py script for an example script that does this. Data preprocessing will execute the first part of the diagram above and create data caches for each specified dataset.

Note: Explicitly preprocessing datasets like this is not necessary; the dataloader will always internally check if there exists a cache for any requested data and will create one if not.

Data Loading

At a minimum, batches of data for training/evaluation/etc can be loaded the following way:

import os
from torch.utils.data import DataLoader
from trajdata import AgentBatch, UnifiedDataset

# See below for a list of already-supported datasets and splits.
dataset = UnifiedDataset(
    desired_data=["nusc_mini"],
    data_dirs={  # Remember to change this to match your filesystem!
        "nusc_mini": "~/datasets/nuScenes"
    },
)

dataloader = DataLoader(
    dataset,
    batch_size=64,
    shuffle=True,
    collate_fn=dataset.get_collate_fn(),
    num_workers=os.cpu_count(), # This can be set to 0 for single-threaded loading, if desired.
)

batch: AgentBatch
for batch in dataloader:
    # Train/evaluate/etc.
    pass

For a more comprehensive example, please see examples/batch_example.py.

For more information on all of the possible UnifiedDataset constructor arguments, please see src/trajdata/dataset.py.

Supported Datasets

Currently, the dataloader supports interfacing with the following datasets:

Adding New Datasets

The code that interfaces the original datasets (dealing with their unique formats) can be found in src/trajdata/dataset_specific.

To add a new dataset, one needs to:

• Create a new folder under src/trajdata/dataset_specific which will contain all the code specific to a particular dataset (e.g., for extracting data into our canonical format). In particular, there must be:
  • An __init__.py file.
  • A file that defines a subclass of RawDataset and implements some of its functions. Reference implementations can be found in the nusc/nusc_dataset.py, lyft/lyft_dataset.py, and eth_ucy_peds/eupeds_dataset.py files.
• Add a subclass of NamedTuple to src/trajdata/dataset_specific/scene_records.py which contains the minimal set of information sufficient to describe a scene. This "scene record" will be used in conjunction with the raw dataset class above and relates to how scenes are stored and efficiently accessed later.
• Add a section to the DATASETS.md file which outlines how users should store the raw dataset locally.
• Add a section to src/trajdata/utils/env_utils.py which allows users to get the raw dataset via its name, and specify if the dataset is a good candidate for parallel processing (e.g., does its native dataset object have a large memory footprint which might not allow it to be loaded in multiple processes, such as nuScenes?) and if it has maps.

Examples

Please see the examples/ folder for more examples, below are just a few demonstrations of core capabilities.

Multiple Datasets

The following will load data from both the nuScenes mini dataset as well as the ETH - University scene from the ETH BIWI Walking Pedestrians dataset.

dataset = UnifiedDataset(
    desired_data=["nusc_mini", "eupeds_eth"],
    data_dirs={  # Remember to change this to match your filesystem!
        "nusc_mini": "~/datasets/nuScenes",
        "eupeds_eth": "~/datasets/eth_ucy_peds"
    },
    desired_dt=0.1, # Please see the note below about common dt!
)

Note: Be careful about loading multiple datasets without an associated desired_dt argument; many datasets do not share the same underlying data annotation frequency. To address this, we've implemented timestep interpolation to a common frequency which will ensure that all batched data shares the same dt. Interpolation can only be performed to integer multiples of the original data annotation frequency. For example, nuScenes' dt=0.5 and the ETH BIWI dataset's dt=0.4 can be interpolated to a common desired_dt=0.1.

Map API

trajdata also provides an API to access the raw vector map information from datasets that provide it.

from pathlib import Path
from trajdata import MapAPI, VectorMap

cache_path = Path("~/.unified_data_cache").expanduser()
map_api = MapAPI(cache_path)

vector_map: VectorMap = map_api.get_map("nusc_mini:boston-seaport")

Simulation Interface

One additional feature of trajdata is that it can be used to initialize simulations from real data and track resulting agent motion, metrics, etc.

At a minimum, a simulation can be initialized and stepped through as follows (also present in examples/simple_sim_example.py):

from typing import Dict # Just for type annotations

import numpy as np

from trajdata import AgentBatch, UnifiedDataset
from trajdata.data_structures.scene_metadata import Scene # Just for type annotations
from trajdata.simulation import SimulationScene

# See below for a list of already-supported datasets and splits.
dataset = UnifiedDataset(
    desired_data=["nusc_mini"],
    data_dirs={  # Remember to change this to match your filesystem!
        "nusc_mini": "~/datasets/nuScenes",
    },
)

desired_scene: Scene = dataset.get_scene(scene_idx=0)
sim_scene = SimulationScene(
    env_name="nusc_mini_sim",
    scene_name="sim_scene",
    scene=desired_scene,
    dataset=dataset,
    init_timestep=0,
    freeze_agents=True,
)

obs: AgentBatch = sim_scene.reset()
for t in range(1, sim_scene.scene.length_timesteps):
    new_xyh_dict: Dict[str, np.ndarray] = dict()

    # Everything inside the forloop just sets
    # agents' next states to their current ones.
    for idx, agent_name in enumerate(obs.agent_name):
        curr_yaw = obs.curr_agent_state[idx, -1]
        curr_pos = obs.curr_agent_state[idx, :2]

        next_state = np.zeros((3,))
        next_state[:2] = curr_pos
        next_state[2] = curr_yaw
        new_xyh_dict[agent_name] = next_state

    obs = sim_scene.step(new_xyh_dict)

examples/sim_example.py contains a more comprehensive example which initializes a simulation from a scene in the nuScenes mini dataset, steps through it by replaying agents' GT motions, and computes metrics based on scene statistics (e.g., displacement error from the original GT data, velocity/acceleration/jerk histograms).

Citation

If you use this software, please cite it as follows:

@Inproceedings{ivanovic2023trajdata,
  author = {Ivanovic, Boris and Song, Guanyu and Gilitschenski, Igor and Pavone, Marco},
  title = {{trajdata}: A Unified Interface to Multiple Human Trajectory Datasets},
  booktitle = {{Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks}},
  month = dec,
  year = {2023},
  address = {New Orleans, USA},
  url = {https://arxiv.org/abs/2307.13924}
}

TODO

• Create a method like finalize() which writes all the batch information to a TFRecord/WebDataset/some other format which is (very) fast to read from for higher epoch training.