PHILADELPHIA, PA — DreamVu said earlier this month it released PRISM, a 270,000-sample multi-view video dataset designed to train and evaluate vision-language models for embodied artificial intelligence tasks.
The dataset was collected across five supermarkets and includes spatial, physical, and action-based reasoning data captured from both worker-worn and 360-degree overhead cameras.
DreamVu said fine-tuning models on PRISM reduced average error rates by 66.6% and cut embodied reasoning errors by a factor of five across 20 capability tests.
The company said existing datasets typically isolate one type of reasoning, while PRISM integrates all three within a single real-world environment.
Annotations were generated using large language model-based reasoning, which DreamVu said improved accuracy compared with template-based labeling, particularly for spatial and causal tasks.
Fourteen of the 20 evaluation benchmarks included in PRISM are not present in previously available public datasets, according to the company.
A data-scaling analysis found that using 60% of the dataset, or about 162,000 samples, achieved 87.7% accuracy, within 1.2 percentage points of results from the full dataset.
Combining egocentric and overhead camera perspectives improved cross-view performance without reducing accuracy on first-person tasks, DreamVu said.
“The core finding is that domain-specific fine-tuning on data covering spatial, physical, and action reasoning together produces gains that general-corpus scaling does not,” said Rajat Aggarwal, co-founder and chief executive officer of DreamVu.
The research paper is available at https://dreamvu.ai/prism, with an arXiv version forthcoming.
A 100,000-sample subset and fine-tuned model weights are available at https://huggingface.co/datasets/DreamVu/PRISM-100K, while the full dataset is offered under commercial license.
DreamVu develops data infrastructure and camera systems used to create training datasets for AI applications in sectors including retail, logistics, healthcare, and manufacturing.
For the latest news on everything happening in Chester County and the surrounding area, be sure to follow MyChesCo on Google News.