VSRToM

VSRToM evaluates 0-order egocentric, 1st-order allocentric and 2nd-order recursive allocentric reasoning in synthetic and real-world scenarios. Both sim and real configurations contain evaluation data in a test split. No train/dev split is provided.

from datasets import load_dataset
sim = load_dataset("Ryngar/VSRToM", "sim", split="test")
real = load_dataset("Ryngar/VSRToM", "real", split="test")

image contains embedded image bytes and a portable filename. question, options, answer, level and task are standardized across configurations. annotation_json contains the remaining research annotations in the original schema, with crop paths and debugging fields removed. Parse it with json.loads. This preserves orientation estimates, synthetic geometry and auxiliary text representations. Ground-truth geometry and auxiliary text are reference annotations and should not be supplied as inputs to image-only/default or DePeR evaluation. For DePeR, use the tool-estimated orientations as specified by the released task implementation.

source_url preserves the original COCO image URL for real-world images. Original record and scene identifiers are for benchmark bookkeeping. Source annotation key order is preserved so person-orientation hint order is reproducible.

Preparation

Absolute local paths and crop filenames are removed. Images are exported as lossless PNG with embedded EXIF, XMP and text metadata removed. Existing PNG pixel streams are retained; JPEGs are encoded from their decoded RGB pixels. The exporter checks that decoded RGB pixels are unchanged. These structural checks do not establish that image contents contain no identifiable people or objectionable material.

Source rights and licensing

This dataset uses component-specific terms, recorded in LICENSE. Newly created VSRToM questions, answer options, answers, and annotations are licensed under CC BY 4.0, to the extent owned by the VSRToM contributors. Real-world photographs retain their original image licenses; COCO annotation licensing does not cover those photographs. Synthetic assets and rendered-image rights require separate verification and are excluded from this annotation grant. The separate code release uses Apache License 2.0.