Augmented ATC Trajectory Dataset

Augmented ATC Trajectory Dataset The Augmented ATC Trajectory Dataset consists of additional labels for who interacts with a robot made on top of the ATC Trajectory Dataset. If you use these labels in conjunction with the ATC Trajectory Dataset, please cite both the paper introducing the augmentation and the following reference for the original data: Y. Kato, T. Kanda, H. Ishiguro, "May I help you? - Design of human-like polite approaching behavior", ACM/IEEE International Conference on Human-Robot Interaction (HRI 2015), pp. 35-42, 2015 Labeling process The original ATC Trajectory Dataset includes scenes where people are tracked as they walk through the ATC shopping mall in Japan. A single 'target' person is labeled as either "interacting" or "not interacting" with the robot at the end of the scene. This augmentation was created by asking two annotators to label all people that they thought interacted with the robot at the end of the scene. The annotator was asked to use the people who were already labeled in the ATC Trajectory dataset as examples of interactants and non-interactants. Annotators were shown videos depicting people moving through the shopping mall. Each person was shown with a black dot and an ID number. The annotated scenes do not depict all scenes in the ATC Trajectory Dataset - only scenes where the robot was stationary were labeled by the annotators. In the original paper introducing this work, only the labels from Annotator 1 are used (found in interacting_ids_annotator_1.json). Using the labels The annotations are in the files interacting_ids_annotator_1.json and interacting_ids_annotator_2.json. Each file contains an ID for the scene and a list of the IDs of all people who interacted with the robot. The original structure of the ATC Trajectory dataset is IntentiontoInteract -> -> dataset_ -> dataset_ -> dataset_ -> dataset_ -> -> -> OtherDistinctiveIntention -> -> -> This folder structure is used to create an ID for each scene. We use an interaction label to simplify the name of the first folder. Scenes from the IntentiontoInteract folder get the interaction label int and scenes from the OtherDistinctiveIntention folder get the interaction label noint. Each scene ID is then constructed as: __dataset_ The structure of the labels is then: { "__dataset_": [, , , , ], ... } In practice, a sample of annotator 1's labels looks like: { "int_20140326_dataset_16203302.csv": [16203302, 16204101], "int_20140326_dataset_15232101.csv": [15232101, 15232200], "int_20140326_dataset_15240800.csv": [15240800, 15245402], ... } Using the labels If you want to use the labels to forecast when an interaction between a person and the robot will occur (as in the original paper), you can find the code to process and use these labels in this [GitLab repo] (https://gitlab.com/interactive-machines/perception/interaction-forecasting-public). Administrivia If you use these labels, please cite both the paper for the original dataset and the paper that published these augmentations. Y. Kato, T. Kanda, H. Ishiguro, "May I help you? - Design of human-like polite approaching behavior", ACM/IEEE International Conference on Human-Robot Interaction (HRI 2015), pp. 35-42, 2015 Thompson, Sydney, Alexander Lew, Yifan Li, Elizabeth Stanish, Alex Huang, Rohan Phanse, and Marynel Vázquez. "Predicting Human Intent to Interact with a Public Robot: The People Approaching Robots Database (PAR-D)." In Proceedings of the 26th International Conference on Multimodal Interaction, pp. 536-545. 2024.