RetailSMV: Exocentric vs. Egocentric Adaptation of Foundation Video World Models in Retail

arXiv 2026 40.4 method, benchmark, application

TLDR

Adapts a foundation video world model to retail scenes, comparing egocentric vs exocentric adaptation using a new synchronized multi-view dataset.

Reasoning

The paper introduces a novel dataset and systematically compares viewpoint adaptation strategies, showing exocentric-only adaptation often outperforms combined. Strengths include rigorous evaluation with multiple metrics and statistical tests; weakness is limited scope to retail domain and no interactive or RL context.

Read-first score

Read-first score 40.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 41.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=dataset,metric

Topical relevance 29%

58.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 244.

Keyword Scores

world model

video world model

world simulator

generative world model

world dynamics prediction

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Introduction of RetailSMV, a synchronized multi-view retail video dataset from store-staff perspective
Systematic comparison of egocentric vs. exocentric adaptation of a foundation video world model to retail scenes
Finding that exocentric-only adaptation outperforms combined adaptation and that adding egocentric data hurts
Identification of the near-horizon prediction window as the regime where adaptation is most beneficial

Methodology

They adapt a pretrained Cosmos3-Nano video diffusion model using Low-Rank Adaptation (LoRA) with three matched configurations (egocentric-only, exocentric-only, combined) on the RetailSMV dataset of 32,105 captioned clips, then evaluate on a 200-clip held-out test set using seven complementary metrics and a strict paired statistical protocol.

Key Results

Exocentric-only adaptation matches or exceeds combined adaptation on six of seven metrics and is significantly better on LPIPS, PSNR, and DreamSim; adding exocentric data to egocentric training helps, while adding egocentric data to exocentric training hurts; the adaptation gap is largest at the shortest rollout time.

Resources

Paper