Multiple-choice questions about how the camera and the objects move in 3D over time in real videos. Questions separate camera motion from object motion: camera translation and turn, object motion after camera compensation, camera–object and object–object distance change, and travel-distance comparison.

This release contains 2,262 QAs on 1,247 videos. These are all QAs that passed the QA-conditioned visual audit, across every motion group. The headline benchmark is its 896-QA / 461-video subset in which both camera motion and scene dynamics are high. The subset is listed in subsets/final_hh_896.txt.

	QAs	Videos
This release (visual-audit keep, all groups)	2,262	1,247
Final benchmark (keep AND camera-high AND dynamics-high)	896	461

Tasks

Every question is asked over measurement frames F0–F8. Directions use the camera axes frozen at the question's start frame: right/down/forward = +x/+y/+z.

Task	`task`	Options	QAs
T1 Camera translation direction	`camera_translation_direction`	6	448
T2 Camera view direction change (yaw/pitch, roll excluded)	`camera_view_direction_change`	4	462
T3 Object world motion direction (camera-compensated)	`object_world_motion_direction`	6	549
T4 Camera–object distance change	`camera_object_distance_change`	2	396
T5 Object–object distance change	`object_pair_distance_change`	2	220
T6 Object travel distance comparison	`object_travel_distance_comparison`	2	187

Motion groups

Each video is labelled high (H) or low (L) on three axes, in the order camera translation / camera rotation / scene dynamics:

Camera translation: cumulative camera path ≥ 0.20 background depths.
Camera rotation: cumulative SO(3) rotation ≥ 40°.
Scene dynamics: camera-compensated foreground flow residual ≥ 0.12 image diagonals and foreground area ≥ 1%.

L does not mean "no motion". The final benchmark is the HH cohort HLH + LHH + HHH.

Group	LLH	HHH	HLH	HHL	LHH	LLL	HLL	LHL
QAs	794	392	314	228	190	173	91	80

Frame protocol

Each released sequence is sampled at 33 uniformly spaced context frames.
Context indices 0, 4, …, 32 are the 9 measurement frames F0–F8 that models see.
Time is normalised to [0, 1] over the released sequence. It does not mean seconds or FPS.
segmentation.json → source_indices maps the 33 context frames to released frame indices.
frame_references in each QA gives the released frame names.

How answers were produced

Answers are automatic geometry references. They are not human labels.

Geometry: VGGT-Omega pose, depth and intrinsics computed on F0–F8.
Objects: full GMOS multi-object masks, with surface points tracked by RAFT-large over the 33 context frames.
Object anchor: the mean of the same surviving tracked surface points. It is not the object's centre of mass.
Units: distances are in sequence background-depth units, not metres.
QA filters: a QA is emitted only when the effect size passes fixed thresholds, a 128-sample spatial-block bootstrap is stable, and tracking and mask quality gates pass.
Object descriptions: generated by a local Qwen model on marked keyframes.
Visual audit: every QA was checked by the same Qwen model, using the question plus the nine marked frames, for visual interpretability. This release keeps only keep decisions.

Records carry quality_status = "provisional_automated_pass" and manual_review = false.

Files

qa/qa.jsonl, qa/qa.json         2,262 QA records (unchanged from the generation run)
qa/visual_audit.jsonl           audit decision per QA
subsets/all_keep_2262.txt       QA IDs in this release
subsets/final_hh_896.txt        QA IDs of the final benchmark
eval/inputs.jsonl               model inputs: prompt, options, frame_paths (F0–F8 with target contours)
eval/reference_answers.jsonl    answer and answer_index per qa_id
eval/protocol.json              evaluation protocol (input policy, scoring)
media/<id>/F0.jpg … F8.jpg      model-input frames; orange = object A, green = object B
cases/<video_id>/
  rgb_0.jpg … rgb_8.jpg         F0–F8 on the geometry grid (unmarked)
  instance_masks.npz            per-object masks for F0–F8, key = object id (e.g. "g1"), uint8 [9, H, W]
  objects.json                  object category / description / identity confidence
  segmentation.json             mask model, scores, context→released frame indices
  selected_record.json          3D anchors per frame, bootstrap samples, camera centres
videos/<source>/<video_id>.zip  all released RGB frames: <video_id>/rgb/00001.jpg … (PNG for SA-V and MVS-Synth)
videos/index.csv                video_id, source, scene_id, num_frames, group, num_qa, zip
summary.json                    counts

Paths inside QA records, such as frame_references[].image_path = "cases/<video_id>/rgb_0.jpg", resolve relative to the dataset root. Masks and rgb_*.jpg share the geometry grid, VGGT-Omega's 416-px preprocessing. Extreme aspect ratios are centre-cropped to [0.5, 2].

QA record fields

Field	Description
`qa_id`, `video_id`	Identifiers
`task`, `group`	Task and 3-letter motion group
`question`, `options`	Question and options; option order is shuffled with a fixed seed
`answer`, `answer_index`	Reference answer
`frame_indices`, `frame_references`	Question window in F-indices; released frame names and normalised times
`object_ids`, `target_mapping`, `target_descriptions`	Targets A/B and their mask ids and descriptions
`evidence`, `thresholds`, `units`	Measured values (e.g. displacement, bootstrap CI) and the fixed thresholds
`answer_source`, `localization_source`, `position_definition`	Provenance

Usage

import json, zipfile
from huggingface_hub import snapshot_download

root = snapshot_download("Ever2after/superdyn-real", repo_type="dataset",
                         allow_patterns=["qa/*", "eval/*", "subsets/*", "media/*"])
qa = [json.loads(l) for l in open(f"{root}/qa/qa.jsonl")]
final = set(open(f"{root}/subsets/final_hh_896.txt").read().split())
benchmark = [q for q in qa if q["qa_id"] in final]          # 896 QAs

inputs = [json.loads(l) for l in open(f"{root}/eval/inputs.jsonl")]
# each input: prompt, options, frame_paths (9 images relative to root)

Full videos are large, about 49 GB in total. Download only what you need, e.g. allow_patterns=["videos/DAVIS/*"].

Evaluation

The reference protocol gives every model the same nine marked frames, the question and the options in the stored order. The model must answer with a single option letter. The prompt is stored in eval/inputs.jsonl. Accuracy uses all QAs in the denominator, so unparsable answers count as wrong. Uniform random accuracy on the final 896 QAs is 28.125%.

Sources and licences

Videos are released RGB frames from DynamicVerse (Apache-2.0 packaging). They were originally collected from the datasets below, whose licence terms still apply to the frames.

Source	Videos
DAVIS	34
MOSE	368
MVS-Synth	107
SA-V	107
VOST	135
UVO	337
YouTube-VIS	159

The QA annotations, masks and evidence were produced by this project. Their licence is to be decided. Use the frames for research only, in line with the most restrictive source licence.

Limitations

Answers are model-estimated. They come from monocular geometry and tracking. Bootstrap checks test tracking stability, not pose or depth error.
Motion groups are labelled per video. They do not guarantee that camera and object motion co-occur inside each question window.
The set is not balanced. Task and source proportions follow what passed the filters.

Downloads last month: 60

Total file size:

50.9 GB