Research
AMD Introduces Semantic Fencing for Video Data Splitting
AMD researchers present a new method for splitting video datasets using embeddings from vision models, reducing semantic leakage in autonomous systems.
Image: AMD
AMD researchers have developed a novel approach for splitting video datasets into training, validation, and test sets by leveraging embeddings from vision foundation models. The method, called Semantic Fencing, addresses the limitations of traditional metadata-based splitting techniques, which often fail to account for the temporal, spatial, and semantic correlations in real-world video streams. Instead of relying on ad hoc metadata rules or random shuffling, the approach uses embeddings to reason about similarity in latent space and construct splits that better reflect true generalization. This technique is particularly relevant for applications like autonomous driving, where continuous data streams exhibit strong correlations and redundancy. The method involves defining bounded semantic regions in latent space and assigning data points to different splits based on their location within these regions. The approach is demonstrated using dashcam sequences from the Zenseact Open Dataset. *Source: [amd](https://rocm.blogs.amd.com/artificial-intelligence/semantic-fencing/README.html)*
Key points
- AMD researchers present a new method for splitting video datasets using embeddings from vision foundation models.
- The approach, called Semantic Fencing, addresses the limitations of traditional metadata-based splitting techniques.
- Semantic Fencing uses embeddings to reason about similarity in latent space and construct splits that better reflect true generalization.
- The method involves defining bounded semantic regions in latent space and assigning data points to different splits based on their location within these regions.
- The approach is demonstrated using dashcam sequences from the Zenseact Open Dataset.