constrained-sdd 0.1.9

A Python package for working with constrained Stanford Drone Dataset.

Constrained Stanford Drone Dataset

This repository contains an annotated version of the Stanford Drone Dataset[1]. We segmented the first 50 images from SDD and drew polygons for buildings, obstacles and offroad. Our trajectories only follow the roads/walking paths.

Installation

Using this dataset is really easy! First, you can install this package using pypi:

 pip install constrained-sdd

then just use the provided class to download the dataset and load it into memory:

import sdd.constrained_sdd as csdd

img_id = 0
folder = "data/sdd"
sdd = csdd.ConstrainedStanfordDroneDataset(0, sdd_data_path=folder, download=True)
train, val, test = sdd.get_dataset()

this creates a dataset with a list of trajectories in train/val/test, e.g. useful for position prediction.

You can also create the trajectory-prediction task via:

train, val, test = sdd.get_trajectory_prediction_dataset(window_size, sampling_rate)

Citation

Leander Kurscheidt, Paolo Morettin, Roberto Sebastiani, Andrea Passerini, Antonio Vergari, A Probabilistic Neuro-symbolic Layer for Algebraic Constraint Satisfaction, arXiv:2503.19466

Examples

An overview over all the images can be seen in analysis/viz_data.ipynb.

Further Details

We annotate the first 50 pictures (scenes in the original dataset) with trajectories of the Stanford Drone Dataset[1], as extracted by P2T[2]. We select all trajectories following the road and delete obvious errors, like projecting a trajectory that slightly touches a building outwards and deleting trajectories that suddenly jump around. We annotate using three classes: Building, Obstacle and Offroad, which forms our constraint-set. Additionally, we have annotated the entrance of a building to differentiate trajectories that enter a building from those who not, but this is unused. All trajectories in our dataset are constraint-abiding.

[1] A. Robicquet, A. Sadeghian, A. Alahi, S. Savarese, Learning Social Etiquette: Human Trajectory Prediction In Crowded Scenes in European Conference on Computer Vision (ECCV), 2016. [2] Nachiket Deo, Mohan M. Trivedi, Trajectory Forecasts in Unknown Environments Conditioned on Grid-Based Plans, 2021

pushing a new version

Pushing a new version is automated via github actions triggered via tags. It is very easy:

1. increment the version number in setup.py
2. git tag -a v0.1.X -m "Release v0.1.X" (adds the tag)
3. git push origin v0.1.8 (pushed the tag)
4. the github action will push a new version