MultiEgoView

This repository contains the real world data from MultiEgoView a dataset generated with EgoSim:

EgoSim: An Egocentric Multi-view Simulator and Real Dataset for Body-worn Cameras during Motion and Activity (NeurIPS 2024)

Dominik Hollidt, Paul Streli, Jiaxi Jiang, Yasaman Haghighi, Changlin Qian, Xintong Liu, Christian Holz
Sensing, Interaction & Perception Lab, Department of Computer Science, ETH Zürich, Switzerland

Dataloading

An example dataloader for the real world data can be found in code/example_dataloader. The dataloader uses the huggingface dataset and can be used for action classification. The huggingface dataset can be wrapped by any frameworks dataset. For fast loading additional preprocessing might be required, e.g. smaller video snippets of equal length or padding.

The synthetic data is stored as individual image frames. While this format is ideal for preserving high-fidelity visual information, it is inefficient in terms of memory usage and access speed, especially when used for video classification or regression tasks in machine learning. To address this, we provide a conversion script (code/dataset_convert.py) that transforms the EgoSimData into more efficient formats such as MP4 or HDF5.