Odia Visual Genome

The Odia Visual Genome is a multimodal dataset comprising aligned textual and visual data, designed to support research in English-Odia multimodal machine translation as well as broader studies in multimodal language processing. The dataset is derived from the Visual Genome corpus, which provides short English image captions paired with corresponding images. For the Odia Visual Genome, we selected a subset of these captions and automatically translated them into Odia, followed by careful manual post-editing. In the post-editing stage, annotators explicitly considered the associated visual context to ensure semantic adequacy and naturalness of the Odia translations. The corpus is partitioned into four subsets. The training set contains approximately 29,000 segments, while the development set and the test set contain 1,000 and 1,600 segments, respectively. Both were taken from Hindi Visual Genome 1.1 where they were created via random sampling from the Visual Genome corpus. In addition, a challenge test set of 1,400 segments was prepared for the WAT2019 Multimodal Translation Task. The challenge test set was constructed to specifically target lexical ambiguity in English captions. Candidate items were identified based on embedding similarity, and ambiguous instances were manually selected where visual information plays a crucial role in disambiguation. Although in many cases surrounding textual context also provides sufficient cues, the inclusion of the image enhances the robustness of disambiguation. Odia Visual Genome was used in WAT 2025 Multimodal Translation Task (https://ufal.mff.cuni.cz/wat2025english-indicmultimodaltranslation). Dataset Formats The dataset contains both textual and visual components. Textual Data. The training, development, and test partitions are distributed as tab-delimited plain-text files. Each file consists of seven columns: Column1 - image_id Column2 - X Column3 - Y Column4 - Width Column5 - Height Column6 - English Text Column7 - Odia Text The bounding-box coordinates (X, Y, Width, Height) specify the rectangular region of the image referenced by the caption. Visual Data. The image collection contains full-resolution images, each identified by the corresponding image_id. The bounding-box metadata enables linking of captions to specific regions within the images. Corpus Statistics Parallel corpus statistics for Odia Visual Genome. Dataset Segments English Words Odia Words ---------------- --------- ---------------- ------------- Train 28930 143134 141652 Dev 998 4922 4912 Test 1595 7854 7734 Challenge Test 1400 8186 8100 ---------------- --------- ---------------- ------------- Total 32923 164096 162398 The word counts are approximate, prior to tokenization.