R4D-Bench

R4D-Bench (CVPR 2026) R4D-Bench is a region-level 4D Video Question Answering benchmark introduced in the CVPR'26 paper: 4D-RGPT: Toward Region-level 4D Understanding via Perceptual Distillation R4D-Bench is the first benchmark to combine region-level prompting with 4D (spatial + temporal) understanding at scale, using dynamic real-world videos. It challenges MLLMs to track, reason about depth, and understand temporal dynamics of specific regions in video. Dataset Summary R4D-Bench contains 1,419 region-prompted multiple-choice VQA pairs built from two existing non-region 4D VQA benchmarks (STI-Bench and VLM4D) via a hybrid automated + human-verified curation pipeline. Each question targets a specific visual region marked with a Set-of-Marks (SoM) token (e.g., ) corresponding to a bounding-box region in the first frame of the video. Questions are multiple-choice with 4–5 options. Splits Split QA Static 388 Dynamic 1,031 Total 1,419 Task Categories Static (3 categories): Category Abbrev. Description 3D Video Grounding VG Retrieve 3D bounding box (dimensions, central point, orientation) of a region Dimension Measurement DM Physical size / distance measurements requiring depth perception Spatial Relationship SR 3D relative positioning between regions or camera Dynamic (6 categories): Category Abbrev. Description Translational T Linear movement direction of a region Rotational R Rotational movement (spin, twist, turn) of a region Counting C Count objects or occurrences of actions False Positive FP Detect when a described event does not actually occur Speed & Acceleration SA Estimate motion dynamics (requires depth + time) Displacement & Path Length DP Measure travel distance of regions across frames Dataset Structure Data Fields { "id": " ", "benchmark": "stibench | vlm4d", "task": " ", "source": "videos/ / .mp4", "time_start": 0.0, "video": "videos/ / .mp4", "som_image": "images/ / .png", // optional "question": " tokens>", "question_raw": " ", "options": ["option_1",