跳转至

正文表格设计

正文暂定四张表,顺序与论文论证一致:先看 Reconstruction,再看完整 Video-to-Physics 流程,然后单独比较 Physics,最后做消融实验。表中暂时不填写没有正式实验来源的数字。

表格 回答的问题
Table 1 Ours Recon 是否比现有方法更准确?
Table 2 从真实视频开始,完整系统是否有效?
Table 3 绕过 Recon 后,Ours Physics 是否仍然更好?
Table 4 改进来自 Recon 对齐、Physics grounding,以及最终方法中的哪些真实组件?

每张正式表都必须在 caption 或对应 manifest 中记录:dataset/protocol、split、training seeds、N scheduled、 CAD/object clusters、failure denominator、matched / adapted / reported-only 状态和 manifest 路径。为保证 9 页正文 可读,这些信息不重复塞进每一列;缺少任一项时,该表仍是空模板,不能填 formal number。

Table 1 · Articulated HOI Reconstruction

D3D-HOI

Object Pose & Dimensions Articulated Motion
Method Reference / provenance Orient. (°) ↓ Loc. (cm) ↓ Dim. (cm) ↓ Rev. (°) ↓ Pri. (cm) ↓
D3D-HOI official Xu et al., arXiv 2021; author-released optimized results
3DADN Qian et al., CVPR 2022; adapted partial-output re-evaluation
Ours full method

D3D-HOI official 和 Ours 使用 D3D-HOI 的 GT-CAD protocol。D3D-HOI 行来自作者发布的 256 个 params.npy,已与官方压缩包逐文件 SHA-256 校验一致,不是我们的复现。3DADN 行只有在 取得逐视频预测并按相同 239-video revolute frame protocol 重评后才保留;否则整行移入附录 reported-only。3DHOI 的 3DADN adaptation 使用 GT motion annotation 做单调片段切分,且 reported State 只统计实际输出 motion parameters 的帧;两项必须与 video/frame coverage 一起披露,不能直接填入 这里的全帧 Rev. 列。

Appendix · 3DHOI legacy protocol(仅作来源对照)

这张表不与上面的 D3D native GT-CAD 主表合并。它完整保留 3DHOI 论文的九项协议,便于解释 3DADN、 CubeOPT 和 D3D-HOI 在原论文中如何比较;所有新数字仍保持空白。

Method Reference / input privilege
3DADN Qian et al., CVPR 2022; 3DHOI adaptation uses GT-motion split; State has subset-frame coverage
CubeOPT Haresh et al., 3DV 2022; our reproduction with official code @ 8965204; GT object + part masks
D3DHOI-GT-CAD Xu et al., arXiv 2021; re-evaluated by 3DHOI; GT object mask + GT CAD
Method Object CD (norm. sq.) ↓ Moving CD (norm. sq.) ↓
3DADN
CubeOPT
D3DHOI-GT-CAD
Method Rotation (°) ↓ Translation (norm.) ↓ Scale (unitless) ↓
3DADN
CubeOPT
D3DHOI-GT-CAD
Method Origin (norm.) ↓ Axis (°) ↓ Direction (°) ↓ State (°) ↓
3DADN
CubeOPT
D3DHOI-GT-CAD

三个紧凑 panel 合计且仅合计原论文九项 error metrics;这里只能填写同一 239-video revolute list、同一 3DHOI evaluator 的数字。CD 是尺度归一 + rigid ICP 后的双向 squared Chamfer;Translation/Origin 也是 normalized-coordinate error,不是 cm;九项均先逐帧、 再逐 video、最后跨 video 平均。CubeOPT 本地结果是我们的 author-code reproduction,不标成 author-released。该表默认进 appendix,不支持 D3D native 主张。

ARCTIC

Hand Reconstruction Object & Articulation HOI Consistency
Method Reference / provenance MPJPE\(_h\) (mm) ↓ MRRPE\(_{r\to l}\) (mm) ↓ MRRPE\(_{r\to o}\) (mm) ↓ AAE (°) ↓ CDev (mm) ↓ MDev (mm) ↓ SR@0.05D (%) ↑
ArcticNet-SF Fan et al., 2023; official baseline
HOPformer Bansal et al., 2026; official code
Ours full method

ARCTIC 只有在三种方法使用相同视角、数据划分和官方评测程序时才保留。ArcticNet-LSTM 仅在论文最终提出 时序论点时加入,否则不占正文行。

Caption: Articulated HOI reconstruction on D3D-HOI and ARCTIC.

Table 2 · End-to-End Video-to-Physics

所有方法使用同一套 frozen Ours Recon 结果,只替换 Physics 方法。这样可以直接比较真实重建误差下的 Physics backend;Ours 一行代表完整系统。唯一 confirmatory primary 在 metric committee 签字前仍标为 Provisional primary。当前推荐 B(normalized-progress outcome)+ full-horizon strict wrapper,不加 contact qualification;若最终签字选择别的候选,只能按同一 签名合同重新聚合所有方法,不能按结果挑阈值。

Task Completion Motion Fidelity ITT audit
Method Provisional primary (%) ↑ Outcome-only (%) ↑ H-MPJPE (mm) ↓ Rev. (rad) ↓ Pri. (m) ↓ Valid coverage (%) ↑ Fail / scheduled ↓
CoDA · Studio adapted
InterMimic · Studio adapted
RePHO · Studio adapted
Ours

RePHO 原生流程使用 rigid-object reconstruction,与本文 articulated-object 设置不同,因此不把它的原生完整 流程直接放入正文排名。正文中的 RePHO 与其他方法一样接收 Ours Recon;原生结果只在附录说明。

Caption: End-to-end video-to-physics results on D3D-HOI. All methods use the same reference reconstructed by our visual stage.

Table 3 · Physics from Captured References

ARCTIC 和 ParaHome 的人体、物体与关节标注直接进入 Physics,不经过 Recon。每个数据集内部,所有方法 使用完全相同的参考动作和评测程序。

Task Completion Motion Fidelity ITT audit
Dataset Method Provisional primary (%) ↑ Outcome-only (%) ↑ H-MPJPE (mm) ↓ Rev. (rad) ↓ Pri. (m) ↓ Valid coverage (%) ↑ Fail / scheduled ↓
ARCTIC CoDA · Studio adapted
InterMimic · Studio adapted
RePHO · Studio adapted
Ours
ParaHome · optional CoDA · Studio adapted
InterMimic · Studio adapted
RePHO · Studio adapted
Ours

ARCTIC captured-reference panel 是优先 controller control;ParaHome 是 optional block。某个数据集的输入转换 没有完成时,整组结果移到附录或删除,不用先导实验数字填正文,也不把两个数据集混成一个 aggregate。

Caption: Physics tracking from captured references. All methods use the same reference within each dataset.

Table 4 · Ablation Studies

External baselines 已在 Tables 1–3 比较;本表只放 Ours 的 paired controls / remove-one ablations,避免把 内部版本伪装成额外方法。

Reconstruction HOI Alignment

Variant Orient. (°) ↓ Loc. (cm) ↓ Rev. (°) ↓ Pri. (cm) ↓ Downstream provisional primary (%) ↑
Ours w/o HOI Align.
Ours

两行使用相同 D3D GT-CAD input、初始化、优化预算和 evaluator;downstream 列使用同一 frozen Physics method。 完整五项 Recon、coverage 与 failure denominator 保留在 A3,本 panel 只留正文论点所需的紧凑列。

Physics Grounding

这部分比较 Physics 前后的结果,说明完整 Physics 阶段是否改善交互和物理合理性。

Variant Object trans. (m) ↓ Object rot. (rad) ↓ Rev. trajectory (rad) ↓ Pri. trajectory (m) ↓
Ours w/o Physics Grounding
Ours
Variant Pairwise contact rate (%) ↑ Contact dist. (mm) ↓ Penetration (mm) ↓ Jerk (m/s³) ↓
Ours w/o Physics Grounding
Ours

这里比较 kinematic reference 与 simulator rollout 的轨迹和物理质量,不声称 Physics 回写或改善 Recon 输出。 没有经过 Physics 执行的结果不能填写 Physics task success。Penetration 只有在碰撞几何和计算方法经过检查后 才保留,否则删除该列;pairwise contact telemetry 未过 replay gate 时同样删除对应列。可用帧 coverage 与 failure denominator 由同一 Table 4 manifest 强制报告在表注/附录,不允许只保留成功片段。

Physics Components

Task Completion Motion Fidelity ITT audit
Variant Provisional primary (%) ↑ Outcome-only (%) ↑ H-MPJPE (mm) ↓ Rev. (rad) ↓ Pri. (m) ↓ Valid coverage (%) ↑ Fail / scheduled ↓
Ours
w/o Residual Action
w/o Contact Conditioning
w/o Articulation Objective

这三行暂时是候选。最终方法中没有独立对应组件时,直接删除该行,不能为了凑表格创造消融。

Caption: Ablation of (a) reconstruction HOI alignment, (b) physics grounding, and (c) implemented physics components. “Provisional primary” remains subject to the signed evaluator contract.

Appendix 空表

正文只保留上面的四张表。下面的表在相应 claim 被激活时填入 appendix;没有完整输入或 GT 时,整张表删除, 不拿 self-consistency 冒充 accuracy。

A1 · Generated-video Reconstruction(无逐帧 GT)

Video protocol / asset source Method Output coverage (%) ↑ Rendered-mask agreement (%) ↑ Valid q duration (%) ↑ Fail / scheduled ↓
VideoGen / Infinigen assets Ours w/o HOI Align.
VideoGen / Infinigen assets Ours
VideoGen / PartNet-Mobility assets Ours w/o HOI Align.
VideoGen / PartNet-Mobility assets Ours

Rendered-mask agreement 是对冻结自动观测的 self-consistency diagnostic,不是 GT reconstruction accuracy。 若 generator 实际导出并冻结逐帧 human/object/q reference,再新增与该 reference 对应的 accuracy 列。

A2 · Generated Video → Recon → Physics

Video protocol / asset source Physics method Provisional primary (%) ↑ Outcome-only (%) ↑ H-MPJPE (mm) ↓ Rev. (rad) ↓ Pri. (m) ↓ Coverage (%) ↑ Fail / scheduled ↓
VideoGen / Infinigen assets CoDA · Studio adapted
InterMimic · Studio adapted
RePHO · Studio adapted
Ours
VideoGen / PartNet-Mobility assets CoDA · Studio adapted
InterMimic · Studio adapted
RePHO · Studio adapted
Ours

若 supporting budget 只允许一个 external Physics baseline,必须在解封结果前按 visible-validation 预注册规则 选择确切方法,并删除其余行;不能在论文中保留匿名占位。

A3 · Reconstruction 消融(D3D-HOI GT-CAD)

Variant Orient. (°) ↓ Loc. (cm) ↓ Dim. (cm) ↓ Rev. (°) ↓ Pri. (cm) ↓ Coverage (%) ↑ Fail / scheduled ↓
Ours w/o HOI Align.
Ours

若最终 Recon 还有第二个可独立删除、且论文声称有效的机制,可在冻结前加一行;否则不为了表格丰富而增加消融。

A4 · Recon noise / controlled corruption

Reference input Corruption level Physics method Provisional primary (%) ↑ Outcome-only (%) ↑ H-MPJPE (mm) ↓ Rev. (rad) ↓ Pri. (m) ↓ Coverage (%) ↑ Fail / scheduled ↓
Captured reference clean Frozen Physics method
Captured reference human noise Frozen Physics method
Captured reference object-pose noise Frozen Physics method
Captured reference articulation noise Frozen Physics method
Captured reference contact timing noise Frozen Physics method

噪声分布、幅度和单位必须在看结果前预注册;revolute 与 prismatic 的正式扰动分别用 rad 和 m,可附带 degree/cm 换算。

A5 · Efficiency 与 failure breakdown

Method Train GPU-hours ↓ Environment transitions ↓ Peak memory (GB) ↓ Eval wall time / case (s) ↓ Crash NaN Timeout Early termination Missing
CoDA · Studio adapted
InterMimic · Studio adapted
RePHO · Studio adapted
Ours

failure 五列填 count,并以相同 N scheduled 为分母;它们不能从 success case 子集中计算。

A6 · Physics diagnostics

Method Object-root trans. (m) ↓ Object-root rot. (rad) ↓ Inactive rev. drift (rad) ↓ Inactive pri. drift (m) ↓ Pairwise contact sustain (%) ↑ Wrong-hand/link (%) ↓ Penetration (mm) ↓ Float/fall (%) ↓
CoDA · Studio adapted
InterMimic · Studio adapted
RePHO · Studio adapted
Ours

Pairwise telemetry、collision geometry 或 support detector 未过 gate 时删除相应列。Revolute 与 prismatic 始终分列,不以一个混合单位均值隐藏失败。

A7 · Physics outcome / wrapper sensitivity

这不是默认必做表。只有 metric committee 认为阈值/定义争议会改变结论时,才加入一个两 panel 紧凑图:

  1. 固定同一签名 wrapper,横轴分别画 A(fixed displacement)、B(normalized progress)、C(endpoint / trajectory tolerance)的预注册阈值,D 报 threshold-sweep AUC;所有方法用同一横轴与 ITT denominator。
  2. 固定同一签名 outcome,只并列 outcome-only、full-horizon strict、valid-duration;contact-qualified 只有 pairwise telemetry gate 通过且论文保留 contact-driven claim 才加入。

Outcome family 与 wrapper 是正交轴,不能任意把 A/B/C/D 各绑到不同 wrapper。图注同时报告 coverage 与 fail/scheduled;若签名 primary 在预注册范围内稳定,本图可省略,只在 artifact 中保留完整 sweep。

A8 · Contact shortcut negative controls(claim-triggered)

Ours condition Provisional primary (%) ↑ Outcome without qualified contact (%) ↓ Pairwise contact sustain (%) ↑ Wrong-hand/link (%) ↓
Normal
Zero contact-intent input
Wrong hand/link replay

这三行是同一方法的预注册 negative controls,不是新 baseline;只有论文保留 contact-driven claim 时才显示。 pairwise telemetry 不能区分正负 replay 时,撤下 contact claim 和本 panel。任何 nonzero hidden object actuation 直接使 rollout 记 Fail,由 common evaluator / failure audit 记录,不再单独占一张论文表。

不预设为论文表的审计证据

  • Native-to-Studio parity 与 train→checkpoint→clean reload→eval 状态在 Physics baseline integration-statusEvidence 维护;审计存在不等于 论文必须展示一张 parity 表。
  • Per-CAD/object-cluster 是 paired bootstrap 的统计单位。只有发现有解释价值的预注册异质性、且样本量足够时, 才从 per_cluster.yaml 生成 optional breakdown;不能事后挑最好/最差 cluster,也不预设一张大表。

Table metadata / protocol index

Panel Dataset / protocol / split Training seeds N scheduled CAD/object clusters Failure denominator Status Experiment protocol
Table 1 · D3D all scheduled case×seed matched / adapted configs/experiments/paper_protocols/recon_real_main.yaml
Table 1 · ARCTIC optional all scheduled case×seed matched, optional configs/experiments/paper_protocols/recon_arctic_optional.yaml
Table 2 · D3D E2E all scheduled case×seed Studio adapted configs/experiments/paper_protocols/pipeline_real_main.yaml
Table 3 · captured Physics all scheduled case×seed Studio adapted configs/experiments/paper_protocols/physics_gt_reference_main.yaml
Table 4 · Recon/Physics ablation all scheduled case×seed ablation / control configs/experiments/paper_protocols/recon_ablation.yaml; physics_ablation.yaml
A1/A2 · Generated supporting all scheduled case×seed supporting configs/experiments/paper_protocols/recon_generated_main.yaml; pipeline_generated_main.yaml
A4 · corruption all scheduled case×seed diagnostic configs/experiments/paper_protocols/recon_corruption.yaml
A5 · efficiency/failure all scheduled case×seed appendix audit configs/experiments/paper_protocols/efficiency_failure.yaml
A6–A8 · claim-triggered diagnostics all scheduled case×seed optional appendix audit configs/experiments/paper_protocols/diagnostics.yaml
ARCTIC optional E2E / ablation all scheduled case×seed optional configs/experiments/paper_protocols/pipeline_arctic_optional.yaml; recon_arctic_ablation_optional.yaml

adapter_parity.yaml 是 readiness/evidence contract;per_cluster.yaml 是结果触发的 statistical breakdown contract。二者不对应预设 paper table,因此不列入上表的 paper-table index。

填表规则

  • 正文只放方法名和科学指标,不增加输入协议、代码状态或失败类型等管理列。
  • 方法不支持的指标写 ;实验尚未完成则留空,不能写 0。
  • 样本数、失败数、有效帧比例、训练预算和更完整的指标放入附录。
  • 表格数字必须能追溯到正式实验结果,不能从开发日志或单个最好案例中选择。