Classifying Objects Using Machine Learning Models for Training
The tech giant Meta has developed a dataset of first-person point of view videos, totaling 12,000 videos, for the purpose of training object classification models. This dataset, which includes videos with varying object distances, camera angles, lighting conditions, and filmed in diverse environments, showcases 200 distinct object categories. Each video was filmed from a first-person perspective, featuring videos with background clutter.
While the dataset is often referenced in research papers or datasets related to egocentric or first-person video data collected by Meta, a direct link or explicit instructions on how to access this specific Meta dataset are not readily available.
Here are some ways you can proceed to access the Meta first-person point of view dataset:
- Check Meta’s official research pages and GitHub repositories associated with their egocentric or first-person video research projects. Large datasets are often linked or referenced there with instructions for access or data usage agreements.
- Review the paper titled “Punching Bag vs. Punching Person: Motion Transferability in First-Person Action Recognition” (2025-07-31), which discusses dataset splits and mentions a GitHub repository for fine/coarse class video data. The GitHub might host part of the dataset or provide download instructions.
- Contact the dataset authors or Meta AI researchers directly when the dataset is not openly distributed. Many datasets require agreeing to terms, registrations, or requests for access.
- Search for related datasets used in Meta’s first-person view research, such as Ego4D or similar egocentric datasets, which might be publicly released for research and can be used for object classification model training.
If you cannot find the Meta dataset, you might also consider using public egocentric datasets like Ego4D or training with synthetic or annotated datasets related to first-person views, as in the synthetic-to-real approaches described in various research papers.
The image associated with this article is credited to Flickr user jose.designs.
- Researchers seeking access to Meta's first-person point of view dataset can check Meta's official research pages and GitHub repositories associated with their egocentric or first-person video research projects, as large datasets are often linked or referenced there with instructions for access or data usage agreements.
- For those who encounter difficulties in locating the dataset directly, it might be beneficial to review the paper titled "Punching Bag vs. Punching Person: Motion Transferability in First-Person Action Recognition" (2025-07-31), as the GitHub repository mentioned in this paper might host part of the dataset or provide download instructions.
