A fundamental challenge in machine translation is generating translations that are semantically accurate, linguistically natural, and culturally appropriate. When relying solely on text, even advanced Large Language Models (LLMs) struggle with lexical ambiguity and lack of context. To address this, we introduce ViVG-10K (Vietnamese Visual Genome), the first manually annotated multimodal dataset for the English–Vietnamese language pair. Based on Visual Genome, ViVG-10K was constructed through a rigorous process involving LLM-generated draft translations and careful post-editing. The dataset comprises 10,013 images with 491,378 English–Vietnamese sentence pairs, cleaned for linguistic quality and visual consistency. We describe the comprehensive construction pipeline, including deduplication, stratified priority sampling, and manual post-editing. Furthermore, we provide a statistical analysis of the dataset and establish benchmarks for both text-only Neural Machine Translation (NMT) and Multimodal Machine Translation (MMT) models.