DynSuperCLEVR

Compositional 4D Dynamic Scenes Understanding with Physics Priors for Video Question Answering ICLR 2025

DynSuperCLEVR is a video question answering (VQA) benchmark for understanding 4D (3D + temporal) object dynamics and physical properties in dynamic scenes. Unlike most video QA datasets that test appearance and 2D motion, DynSuperCLEVR targets the physical concepts that require an explicit 4D scene understanding: velocity, acceleration, and collisions. It asks models not only to describe what happened, but to predict what will happen next and to reason about counterfactual "what-if" scenarios under altered physical conditions.

🌐 Project page	https://xingruiwang.github.io/projects/DynSuperCLEVR/
📄 Paper (arXiv)	https://arxiv.org/abs/2406.00622
📄 Paper (OpenReview)	https://openreview.net/pdf?id=6Vx28LSR7f
🧠 Model code (NS-4DPhysics)	https://github.com/XingruiWang/NS-4DPhysics
🛠️ Dataset generation code	https://github.com/XingruiWang/DynSuperCLEVR
📊 Poster	https://iclr.cc/virtual/2025/poster/30879

---

Highlights

• Physics-grounded VQA. Questions are built around three physical concepts — velocity, acceleration, and collision — that cannot be answered from a single frame and require reasoning over object dynamics through time.
• Three reasoning types. Factual (what is happening), Predictive (what will happen next), and Counterfactual (what would happen under a different initial condition, e.g. "if the engine were turned on" / "if it were static").
• Full 4D annotations. Every scene ships with complete physical ground truth: per-frame 3D positions, linear & angular velocities, camera parameters (intrinsics K and extrinsics R), instance segmentation, and collision events (contact normal, force, frame, image position).
• Rendered with a physics engine. Scenes are generated by extending the Kubric pipeline (built on PyBullet) over the SuperCLEVR asset set, with added support for acceleration dynamics.

---

Dataset statistics

• 1,100 videos (rigid-body simulations), 120 rendered frames each (RGB, instance masks, segmentation).
• 12,752 questions across three reasoning types and two splits.

Reasoning type	Train	Val	Total
Factual	7,850	791	8,641
Predictive	2,750	274	3,024
Counterfactual	989	98	1,087
Total	11,589	1,163	12,752

Answers are a mix of binary (True/False), direction (left, right, front, back, up, down), and attribute (color / object) labels, following the CLEVR-style compositional program format.

---

Repository structure

DynSuperCLEVR/
├── questions/
│   ├── train/
│   │   ├── questions_physics_factual.json
│   │   ├── questions_physics_predictive.json
│   │   └── questions_physics_counterfactual.json
│   └── val/
│       ├── questions_physics_factual.json
│       ├── questions_physics_predictive.json
│       └── questions_physics_counterfactual.json
├── frames_upload_chunks/
│   └── frames_chunk_000.tar ... frames_chunk_023.tar   # all video frames (~325 GB), tarred
└── videos/frames/<scene>/                              # sample uncompressed scenes
    ├── rgba_00000.png ... rgba_00119.png               # RGB frames
    ├── segmentation_00000.png ...                      # semantic segmentation
    ├── mask_00000.png ...                              # instance masks
    ├── metadata.json                                   # camera + per-instance 4D state
    └── events.json                                     # collision events

Question schema

Each question file is a dict with a questions list; every entry contains:

Field	Description
question	Natural-language question
answer	Ground-truth answer
program	Compositional functional program (CLEVR-style) used to derive the answer
image_filename / image	The scene this question refers to (e.g. super_clevr_0)
image_index, question_index, question_family_index	Indices
template_filename	Question template used to generate the question
frame_length	Number of frames the question is grounded over
split	train / val

The Dataset Viewer (and the questions config loadable via datasets) is served from flattened Parquet files under viewer/ (train.parquet, val.parquet), which merge the three reasoning types and add a reasoning_type column. There, the program field is stored as a JSON string for a stable schema; the canonical nested JSON lives under questions/.

Per-scene physical ground truth (metadata.json)

• camera — intrinsics K, extrinsics R, projection of 3D → image.
• instances — for each object: per-frame positions, velocities, angular_velocities, plus shape / color / material attributes.
• metadata — scene-level config: gravity, frame_rate, num_frames, num_instances, background HDRI, resolution.
• args — full generation arguments (floor friction/restitution, camera mode, etc.).

Collision events (events.json)

• collisions — list of contacts with contact_normal, force, frame (sub-frame precise), and image_position.

---

Usage

The frame data is distributed as tar chunks. Download and extract:

# Download the whole dataset
huggingface-cli download RyanWW/DynSuperCLEVR --repo-type dataset --local-dir ./DynSuperCLEVR

# Or just the questions (small) to get started
huggingface-cli download RyanWW/DynSuperCLEVR --repo-type dataset \
    --include "questions/**" --local-dir ./DynSuperCLEVR

# Reassemble the frames
cd DynSuperCLEVR
for f in frames_upload_chunks/frames_chunk_*.tar; do tar -xf "$f"; done

import json

qs = json.load(open("DynSuperCLEVR/questions/val/questions_physics_factual.json"))["questions"]
print(len(qs), "questions")
print(qs[0]["question"], "->", qs[0]["answer"])
# e.g. "Which direction is the red cruiser moving towards at the beginning?" -> "back"

---

Example questions

Type	Question	Answer
Factual	Is the red mountain moving fast at the beginning?	False
Factual	Which direction is the red cruiser moving towards at the beginning?	back
Predictive	Will the purple minivan collide with the suv next?	True
Counterfactual	Will the green fighter collide with the mountain if it turns the engine on?	True

---

Method: NS-4DPhysics

The accompanying model, NS-4DPhysics, is a neural-symbolic approach that builds an explicit 4D scene representation and integrates physics priors:

1. 3D scene parser — a neural mesh model that recovers object pose and state by render-and-compare.
2. Question parser — translates language into a symbolic program.
3. Program executor — runs the program over the 4D representation, using a PyBullet physics engine to roll out dynamics for predictive / counterfactual reasoning.

On DynSuperCLEVR, NS-4DPhysics reaches 82.64% average accuracy, substantially outperforming PO3D-VQA (62.93%), PLLaVA (59.24%), and GPT-4o with reasoning (56.06%), highlighting the gap between current video-language models and explicit physical scene understanding.

---

Citation

@inproceedings{wang2025compositional,
  title     = {Compositional 4D Dynamic Scenes Understanding with Physics Priors for Video Question Answering},
  author    = {Wang, Xingrui and Ma, Wufei and Wang, Angtian and Chen, Shuo and Kortylewski, Adam and Yuille, Alan},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2025},
  eprint    = {2406.00622},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url       = {https://arxiv.org/abs/2406.00622}
}

License

Released under the Apache 2.0 license. DynSuperCLEVR builds on SuperCLEVR and the Kubric generation pipeline; please also respect their respective licenses.

Authors

Xingrui Wang, Wufei Ma, Angtian Wang (Johns Hopkins University); Shuo Chen (Tsinghua University); Adam Kortylewski (Max Planck Institute for Informatics / University of Freiburg); Alan Yuille (Johns Hopkins University).