pith. sign in

arxiv: 2510.02311 · v2 · submitted 2025-10-02 · 💻 cs.CV · cs.LG

Inferring Dynamic Physical Properties from Video Foundation Models

Guanqi Zhan , Xianzheng Ma , Weidi Xie , Andrew Zisserman This is my paper

Pith reviewed 2026-05-18 10:04 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords video foundation modelsphysical property inferenceelasticityviscositydynamic frictiongenerative video modelsself-supervised video learningmultimodal large language models
0
0 comments X

The pith

Pre-trained video foundation models can infer dynamic physical properties like elasticity, viscosity, and friction from video footage using simple readouts or prompts.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to show that video foundation models pre-trained either generatively or in a self-supervised way can be adapted to estimate physical properties that unfold over time, such as the elasticity of a bouncing object, the viscosity of a flowing liquid, and the dynamic friction of a sliding object. It creates new datasets with synthetic training and test splits plus a real-world evaluation split to measure how well this works outside controlled conditions. The authors compare three approaches: an oracle that feeds classical computer vision cues directly into the estimator, a lightweight readout that adds a visual prompt and trainable vector for cross-attention on frozen foundation models, and various prompting strategies applied to multimodal large language models. They report that the generative and self-supervised foundation models reach roughly the same level of accuracy, both behind the oracle but ahead of current MLLMs, whose results improve when prompts are chosen carefully. A sympathetic reader would care because this suggests that large video models already encode enough temporal and physical structure to support material-property inference without building physics simulators from scratch.

Core claim

The central claim is that a video foundation model trained in a generative manner such as DynamiCrafter or in a self-supervised manner such as V-JEPA-2 can be equipped with a simple readout mechanism or prompting strategy to infer dynamic physical properties from videos at a level generally similar to each other though still below an oracle that supplies classical computer vision cues, while multimodal large language models currently lag but can be improved by suitable prompting.

What carries the argument

A readout mechanism that applies a visual prompt together with a trainable prompt vector for cross-attention inside a frozen pre-trained video foundation model, or equivalent prompt strategies inside MLLMs, to regress scalar values for elasticity, viscosity, or dynamic friction.

If this is right

  • Generative and self-supervised video pre-training objectives capture comparable temporal information relevant to physical property inference.
  • A lightweight readout added to frozen foundation models is sufficient to extract dynamic properties without full fine-tuning.
  • Prompt engineering offers a practical route to raise MLLM performance on the same physical-inference tasks.
  • The new synthetic-plus-real datasets provide a standardized testbed for measuring generalization from controlled to uncontrolled video.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the same readout works for elasticity, viscosity and friction, it may also support inference of other time-dependent properties such as thermal conductivity or restitution coefficients when suitable video data exists.
  • Successful inference of these properties from video could be chained to improve downstream video prediction models that must respect physical consistency.
  • The gap between oracle and foundation-model performance on real videos highlights a concrete target for future self-supervised objectives that explicitly reward physical plausibility.

Load-bearing premise

The real split collected for each dataset serves as a faithful proxy for uncontrolled real-world video conditions so that accuracy measured on it indicates how well the methods will generalize to arbitrary physical videos.

What would settle it

A result on the real test split in which the readout from either DynamiCrafter or V-JEPA-2 shows no statistically significant advantage over the best MLLM prompting baseline for viscosity estimation, or in which both fall far below the oracle method on elasticity estimation.

Figures

Figures reproduced from arXiv: 2510.02311 by Andrew Zisserman, Guanqi Zhan, Weidi Xie, Xianzheng Ma.

Figure 1
Figure 1. Figure 1: Examples of the PhysVid dataset. Each row shows a different property, and each column shows three frames from video samples in the synthetic sets (train, test-1, and test-2) and the real test-3 set. The train and test-1 sets are from the same distribution. In test-2 parameters, such as lighting, viewpoint and color, differ from those in test-1. captured from the same viewpoint, and the model must determine… view at source ↗
Figure 2
Figure 2. Figure 2: Oracle methods for physical properties. The objective in each case is to extract a measurement from the sequence that can directly be used to predict the property. For elasticity, we extract the centroid trajectory from segmentation masks, and then normalize the y-coordinates into 0-1; the ratio of bouncing to dropping height over the sequence indicates the elasticity. For viscosity, we calculate the area … view at source ↗
Figure 3
Figure 3. Figure 3: Architectures for dynamic physical property prediction. Left: video generative model as backbone; Middle: video self-supervised model as backbone; Right: multimodal large language model (MLLM). For the pre-trained video diffusion model (U-Net, left) and the pre-trained self￾supervised model (ViT, middle), the representations are kept frozen, and a ‘visual prompt’ learns to infer the physical properties. Fo… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results. Top Left: An example for elasticity absolute value prediction; Bottom Left: An example for friction relative value comparison. For each example, the original input video is shown on the left. A static red circle is overlaid in the center to highlight the full trajectory of the object on every frame, shown in the middle. Model predictions are shown on the right, including results from t… view at source ↗
Figure 5
Figure 5. Figure 5: Objects and surfaces in the friction real dataset. Top: Objects used for friction real dataset collection; Bottom: Surfaces used for friction real dataset collection. B.3 DEVICES FOR REAL DATASET COLLECTION [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Devices used to collect real datasets. Left: The funnel used in the collection of the viscosity real dataset; Middle-left: The funnel holder used in the collection of the viscosity real dataset; Middle-right: The spring dynamometer used to measure the ground truth dynamic friction coefficient in the collection of the friction real dataset; Right: The slope used to give the objects an initial velocity on th… view at source ↗
Figure 7
Figure 7. Figure 7: shows the visual input to the MLLM; [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Example of baseline prompt for the absolute formulation. The example is on Gemini for the elasticity property. The initial state of object motion is incorrectly recognized from the beginning. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example of oracle estimation teaching for the absolute formulation. The example is on Gemini for the elasticity property. Although the model strictly follows the oracle’s step-by-step guidance, an incorrect identification of the peak in the third step leads to a significantly inaccurate final prediction. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Example of few-shot examples for the absolute formulation. The example is on Gemini for the elasticity property. The ground truth examples provided in the few-shot setting serve as effective calibration signals, leading to notably improved performance. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Example of frame index provided for the absolute formulation. The example is on Gemini for the elasticity property. Providing frame indices helps the model better interpret the motion process. However, estimating the final value based solely on this information remains chal￾lenging. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: shows the visual input to the MLLM; [PITH_FULL_IMAGE:figures/full_fig_p026_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Example of baseline prompt for the relative formulation. The example is on Gemini for the elasticity property. The baseline model exhibits reasonable performance. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Example of oracle estimation teaching for the relative formulation. The example is on Gemini for the elasticity property. The oracle strategy promotes qualitative analysis (e.g., comparing motion or relative magnitudes) without forcing exact calculations. This flexible reasoning process leads to more reliable outputs. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Example of few-shot examples for the relative formulation. The example is on Gemini for the elasticity property. The relative task is simpler—determining which of two instances has a greater physical value—without requiring exact numerical estimates. Here, few-shot examples tend to degrade performance, often encouraging shortcut responses that reduce interpretability and stability. 29 [PITH_FULL_IMAGE:fi… view at source ↗
Figure 16
Figure 16. Figure 16: Example of frame index provided for the relative formulation. The example is on Gemini for the elasticity property. Providing the frame indices enhances the model’s understanding of temporal dynamics, thereby resulting in more effective comparative reasoning. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Example of black frames in between for the relative formulation. The example is on Gemini for the elasticity property. Concatenating both videos with black frames in between enables the model to better perform relative comparisons, likely by making inter-video relationships more explicit. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: shows the scatter plots of oracle estimation on different test splits of the three dynamic physical properties. Elasticity Viscosity Friction Test-1 Test-2 Test-3 [PITH_FULL_IMAGE:figures/full_fig_p032_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: shows the scatter plots of DynamiCrafter on different test splits of the three dynamic physical properties. Elasticity Viscosity Friction Test-1 Test-2 Test-3 [PITH_FULL_IMAGE:figures/full_fig_p033_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: shows the scatter plots of V-JEPA-2 on different test splits of the three dynamic physical properties. Elasticity Viscosity Friction Test-1 Test-2 Test-3 [PITH_FULL_IMAGE:figures/full_fig_p034_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: shows the scatter plots of Qwen2.5VL-max on different test splits of the three dynamic physical properties. For test-1 and test-2, due to the limitation of resources, a random subset of 100 samples are used. Elasticity Viscosity Friction [PITH_FULL_IMAGE:figures/full_fig_p035_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: shows the scatter plots of GPT-4o on different test splits of the three dynamic physical properties. For test-1 and test-2, due to the limitation of resources, a random subset of 100 samples are used. Elasticity Viscosity Friction [PITH_FULL_IMAGE:figures/full_fig_p036_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: shows the scatter plots of Gemini-2.5-pro on different test splits of the three dynamic physical properties. For test-1 and test-2, due to the limitation of resources, a random subset of 100 samples are used. Elasticity Viscosity Friction [PITH_FULL_IMAGE:figures/full_fig_p037_23.png] view at source ↗
read the original abstract

We study the task of predicting dynamic physical properties from videos. More specifically, we consider physical properties that require temporal information to be inferred: elasticity of a bouncing object, viscosity of a flowing liquid, and dynamic friction of an object sliding on a surface. To this end, we make the following contributions: (i) We collect a new video dataset for each physical property, consisting of synthetic training and testing splits, as well as a real split for real world evaluation. (ii) We explore three ways to infer the physical property from videos: (a) an oracle method where we supply the visual cues that intrinsically reflect the property using classical computer vision techniques; (b) a simple read out mechanism using a visual prompt and trainable prompt vector for cross-attention on pre-trained video generative and self-supervised models; and (c) prompt strategies for Multi-modal Large Language Models (MLLMs). (iii) We show that a video foundation model trained in a generative (DynamiCrafter) or trained in a self-supervised manner (V-JEPA-2) achieve a generally similar performance, though behind that of the oracle, and that MLLMs are currently inferior to the other models, though their performance can be improved through suitable prompting. The dataset, model, and code are available at https://www.robots.ox.ac.uk/~vgg/research/idpp/.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper studies the task of inferring dynamic physical properties (elasticity of bouncing objects, viscosity of flowing liquids, dynamic friction of sliding objects) from video. It contributes new datasets with synthetic train/test splits plus a real split, compares an oracle using classical CV cues, a readout mechanism on pre-trained video models (DynamiCrafter generative and V-JEPA-2 self-supervised) via visual prompts, and prompting strategies for MLLMs. The central claim is that the two foundation models achieve generally similar performance (behind the oracle but ahead of MLLMs, whose results improve with suitable prompting).

Significance. If the real-world evaluation is valid, the work demonstrates that generative and self-supervised video foundation models implicitly encode dynamic physical properties without task-specific training, positioning them as competitive with classical CV oracles for this inference task. The public release of the dataset, model, and code is a clear strength for reproducibility.

major comments (2)
  1. [Dataset collection] Dataset section (contribution i and real-split description): The manuscript presents the real split as the key test of generalization to uncontrolled physical videos, yet supplies no details on scene diversity, camera motion, lighting variation, material range, or independent ground-truth acquisition (e.g., lab-measured restitution coefficients or viscosity versus visual estimation). This assumption is load-bearing for the claim that measured performance indicates usable real-world inference.
  2. [Results and evaluation] Results section: The abstract and evaluation claim that DynamiCrafter and V-JEPA-2 achieve 'generally similar performance' and that MLLMs are inferior (but improvable), but the provided text contains no quantitative metrics, error bars, dataset statistics, or statistical tests supporting the performance ordering or the gap to the oracle; this leaves the central empirical claims difficult to assess.
minor comments (2)
  1. [Abstract] Abstract: The high-level performance ordering would be clearer if the specific metrics (e.g., MAE, correlation) were named even briefly.
  2. Notation: Ensure consistent use of 'dynamic friction' versus 'friction coefficient' across text and figures.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which highlight important aspects for strengthening the presentation of our work on inferring dynamic physical properties from video foundation models. We address each major comment below and outline the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: [Dataset collection] Dataset section (contribution i and real-split description): The manuscript presents the real split as the key test of generalization to uncontrolled physical videos, yet supplies no details on scene diversity, camera motion, lighting variation, material range, or independent ground-truth acquisition (e.g., lab-measured restitution coefficients or viscosity versus visual estimation). This assumption is load-bearing for the claim that measured performance indicates usable real-world inference.

    Authors: We agree that more detailed documentation of the real split is essential to substantiate the generalization claims. In the revised manuscript, we will expand the dataset section with explicit descriptions of scene diversity (e.g., indoor/outdoor environments and object types), camera motion variations, lighting conditions, and material ranges. Regarding ground-truth acquisition, labels for the real videos were obtained via multi-annotator visual estimation following standardized protocols for each property (e.g., counting bounces for elasticity), with inter-annotator agreement reported; we acknowledge that independent lab measurements would provide stronger validation but were not available due to practical constraints in collecting uncontrolled real-world footage. We will clarify this distinction and its implications for the results. revision: yes

  2. Referee: [Results and evaluation] Results section: The abstract and evaluation claim that DynamiCrafter and V-JEPA-2 achieve 'generally similar performance' and that MLLMs are inferior (but improvable), but the provided text contains no quantitative metrics, error bars, dataset statistics, or statistical tests supporting the performance ordering or the gap to the oracle; this leaves the central empirical claims difficult to assess.

    Authors: We apologize for the lack of explicit numerical summaries in the narrative text. The full quantitative results—including mean performance metrics with standard deviations (error bars), dataset statistics (e.g., split sizes and property distributions), and comparisons across methods—are reported in Tables 1–3 and visualized in Figures 2–5 of the results section. In the revision, we will insert direct textual references to these tables and figures, along with a concise summary paragraph highlighting key values and the observed ordering (foundation models behind oracle but ahead of MLLMs). While we did not include formal statistical significance tests, we can add them in the revision if the referee deems it necessary; the current evidence for 'generally similar performance' rests on the consistent trends across multiple properties and splits shown in the tables. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical comparisons on newly collected datasets

full rationale

The paper introduces new video datasets with synthetic train/test splits and a real split, then runs controlled experiments comparing an oracle (classical CV cues), readout heads on pre-trained video foundation models (DynamiCrafter, V-JEPA-2), and prompted MLLMs. No equations, derivations, or first-principles claims are present. Performance numbers are measured on held-out data rather than being fitted parameters renamed as predictions. No self-citation chains or ansatzes are invoked to justify core results. The evaluation is self-contained against external benchmarks (new data collection and model comparisons).

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work rests on the standard computer-vision premise that video appearance encodes dynamic physical information, with no free parameters, new entities, or ad-hoc axioms introduced in the abstract.

axioms (1)
  • domain assumption Video footage contains sufficient visual information to infer dynamic physical properties such as elasticity, viscosity, and friction.
    This premise is required for the inference task to be solvable from video input alone.

pith-pipeline@v0.9.0 · 5781 in / 1259 out tokens · 54939 ms · 2026-05-18T10:04:51.686896+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. PhysInOne: Visual Physics Learning and Reasoning in One Suite

    cs.CV 2026-04 unverdicted novelty 8.0

    PhysInOne is a new dataset of 2 million videos across 153,810 dynamic 3D scenes covering 71 physical phenomena, shown to improve AI performance on physics-aware video generation, prediction, property estimation, and m...

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages · cited by 1 Pith paper · 6 internal anchors

  1. [1]

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

    Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Matthew Muckley, Am- mar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, et al. V-jepa 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Fitvid: Overfitting in pixel-level video prediction.arXiv preprint arXiv:2106.13195,

    Mohammad Babaeizadeh, Mohammad Taghi Saffar, Suraj Nair, Sergey Levine, Chelsea Finn, and Dumitru Erhan. Fitvid: Overfitting in pixel-level video prediction.arXiv preprint arXiv:2106.13195,

    work page arXiv

  3. [3]

    Bear, Elias Wang, Damian Mrowca, Felix J

    Daniel M Bear, Elias Wang, Damian Mrowca, Felix J Binder, Hsiao-Yu Fish Tung, RT Pramod, Cameron Holdaway, Sirui Tao, Kevin Smith, Fan-Yun Sun, et al. Physion: Evaluating physical prediction from vision in humans and machines.arXiv preprint arXiv:2106.08261,

    work page arXiv

  4. [4]

    Intphys 2: Benchmarking intuitive physics understanding in complex synthetic environ- ments.arXiv preprint arXiv:2506.09849,

    Florian Bordes, Quentin Garrido, Justine T Kao, Adina Williams, Michael Rabbat, and Emmanuel Dupoux. Intphys 2: Benchmarking intuitive physics understanding in complex synthetic environ- ments.arXiv preprint arXiv:2506.09849,

    work page arXiv

  5. [5]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capa- bilities.arXiv preprint arXiv:2507.06261,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, et al. Qwen2. 5-coder technical report.arXiv preprint arXiv:2409.12186,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    GPT-4o System Card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Sora: A Review on Background, Technology, Limitations, and Opportunities of Large Vision Models

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. InEuropean Conference on Computer Vision (ECCV), 2024a. Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Hua...

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    Grounding dino 1.5: Advance the "edge" of open-set object detection, 2024

    11 Tianhe Ren, Qing Jiang, Shilong Liu, Zhaoyang Zeng, Wenlong Liu, Han Gao, Hongjie Huang, Zhengyu Ma, Xiaoke Jiang, Yihao Chen, Yuda Xiong, Hao Zhang, Feng Li, Peijun Tang, Kent Yu, and Lei Zhang. Grounding dino 1.5: Advance the ”edge” of open-set object detection.arXiv preprint arXiv:2405.10300, 2024a. Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Ku...

    work page arXiv

  10. [10]

    Phyx: Does your model have the” wits” for physical reasoning?arXiv preprint arXiv:2505.15929,

    Hui Shen, Taiqiang Wu, Qi Han, Yunta Hsieh, Jizhou Wang, Yuyue Zhang, Yuxin Cheng, Zijian Hao, Yuansheng Ni, Xin Wang, et al. Phyx: Does your model have the” wits” for physical reasoning?arXiv preprint arXiv:2505.15929,

    work page arXiv

  11. [11]

    Masked conditional video diffusion for prediction, generation, and interpolation.arXiv preprint arXiv:2205.09853,

    Vikram V oleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Masked conditional video diffusion for prediction, generation, and interpolation.arXiv preprint arXiv:2205.09853,

    work page arXiv

  12. [12]

    Neural Material: Learning Elastic Constitutive Material and Damping Models from Sparse Data

    Bin Wang, Paul Kry, Yuanmin Deng, Uri Ascher, Hui Huang, and Baoquan Chen. Neural mate- rial: Learning elastic constitutive material and damping models from sparse data.arXiv preprint arXiv:1808.04931,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    A.3 DETAILS OFVIDEOINPUT We uniformly sample 16 frames per video as input to all the models for fair comparison. The 16 frames are uniformly sampled so that the physics process we want to study is properly reflected with the sampled 16 frames,e.g.,the dropping and bouncing of the ball, the expansion of the liquid, and the slowing-down sliding process of t...

    work page 2025