Researchers found that the camera angle used to train an AI video model matters more than the volume of footage.
A team studying how to adapt general-purpose AI video models to retail environments built a dataset called RetailSMV: 32,105 captioned clips from five supermarkets, shot from two angles simultaneously. One angle was egocentric — a wearable camera on store staff performing tasks like stocking shelves, scanning items, and managing supply carts. The other was exocentric — fixed overhead or third-person cameras watching the same activity. They then fine-tuned three versions of a compact video model called Cosmos3-Nano using Low-Rank Adaptation, a parameter-efficient technique that updates only a small subset of a model's weights. The three versions trained on egocentric footage only, exocentric only, and both combined.
The finding cuts against the intuition that more data wins. The exocentric-only model, trained on roughly 16,000 clips, matched or beat the combined model on six of seven evaluation metrics — and was statistically superior on three of them — despite the combined model having access to twice the footage. More striking: adding egocentric footage to an exocentric model made performance worse, while the reverse held. The gains were sharpest at short prediction horizons, suggesting fixed-angle video is especially valuable when a model needs to predict what happens next in the immediate future.
The retail AI space is crowded with customer-facing computer vision systems, but this research focuses on the store-operations side — staff workflows rather than shopper behavior. The takeaway for anyone building video AI for warehouses, factories, or retail floors is blunt: pick your camera placement carefully, because more footage from the wrong angle can actively hurt you.