Facebook PE-Video with Slovenian Annotations
Opis
This service provides a localized version of the large-scale PE-Video dataset (originally released by Meta AI). The dataset comprises 1 million motion-centered videos across diverse domains including food preparation, industrial work, and outdoor scenes, with 120,000 of them being human-checked. This service focuses on providing high-quality, linguistically accurate Slovenian annotations. It bridges the gap for low-resource languages in the multimodal AI space, enabling organizations to develop and evaluate models that can see and describe actions in Slovenian, reducing the dependency on English-centric vision-language models. The dataset is constructed using a hybrid machine-translation and human-in-the-loop pipeline. Initial translations of the complex motion-descriptive captions are performed using state-of-the-art neural machine translation and large language models optimized for the Slovenian linguistic structure. These are then subjected to verification by native speakers to ensure temporal logic and action-specific terminology remain accurate to the visual context. The dataset is structured for compatibility with the Perception Encoder (PE) architecture, supporting frame-level alignment and temporal grounding. It utilizes standard formats like WebDataset and JSON-LD to ensure seamless integration into existing deep-learning frameworks such as PyTorch.
Pričakovani rezultati
Slovenian researchers and SMEs can build specialized AI applications such as automated video monitoring for local industry or educational tools that interact naturally with users in their native language. The dataset serves as a foundational benchmark for Slovenian Video Question Answering (VQA), automated video captioning, and cross-modal retrieval.
Kartica storitve
| Kategorija | Infrastructure and data-related services |
| Domena | Multimodal AI, Low-resource Language Data |
| Vertikala | Cross-domain |
| Ciljni uporabniki | Researchers, Software SMEs |
| Tip storitve | Dataset |
| Rezultat | Comprehensive set of Slovenian language annotations for the 120k human-verified subset of the PE-Video collection, including detailed video descriptions, action keywords, and temporal segments. |
| Zahteve | |
| Omejitve | |
| Na voljo od | Q4/2026 |
| Licenca | |
| Slika | |
| Oznaka slike |