Movie archives still rely on manual cataloging and sparse metadata, limiting fine-grained retrieval, relationship tracing, and reuse under privacy-constrained edge settings. We propose EdgeCineTag-KG, an edge framework using a single video foundation encoder and knowledge-constrained multi-label learning to produce consistent labels and build a queryable movie-archive knowledge graph. The objective jointly models label co-occurrence, mutual exclusion, hierarchy, and temporal consistency to reduce semantic contradictions and label jitter. For deployment, an uncertainty-driven adaptive computation strategy meets real-time constraints with controlled quality loss. Across MovieNet, Condensed Movies, Trailers12k, MMTF-14K, and TVQA, performance improves from 47.8 to 55.6 mAP and from 38.2 to 44.9 Macro-F1 on MovieNet, from 42.1 to 49.3 mAP on Condensed Movies, and from 71.2 to 75.4 mAP on Trailers12k. Knowledge graph quality also improves, with rule violation rate dropping from 6.8% to 2.4% and link prediction MRR rising from 0.248 to 0.312. Under INT8 adaptive inference, the system reaches 5.3 Clip-FPS, 182 ms P95 latency, and 1.9 GB peak memory. This combination improves consistency and retrieval usability without relying on multiple stacked foundation models. These results support reliable, interpretable, and edge-deployable movie archive understanding.
Qi et al. (Sat,) studied this question.