acceptodds
Under review as a conference paper at ICLR 2027

CAHI4D: Cross-Age Human Interaction 4D Reconstruction from Monocular Video

Abstract

Age-dependent morphology exacerbates scale–depth ambiguity in monocular multi-person reconstruction, while adult-centric priors and occlusion of interaction cues further hinder cross-age interaction recovery. We present Cross-Age Human Interaction 4D Reconstruction (CAHI4D), which uses a reliable reference and validated interactions for geometric anchoring before fitting age-dependent shape and refining interaction motion. A vision–language model (VLM) proposes support and contact hypotheses from visual context; geometric and temporal checks then determine which relations can constrain reconstruction. Depth- and Interaction-Guided Geometric Anchoring (DIGA) uses a reliable scene-supported reference to align the shared scene scale, and then places a geometrically underconstrained participant through validated interactions and corrects body scale from the induced projection change. With fixed placement and overall body scale, shape fitting adapts age-dependent proportions. On the fixed geometry, Kinematic-Topology-Guided Interaction Correction (KTIC) maps directed surface corrections to participant-specific kinematic chains and reconciles corrections sharing joint degrees of freedom; contact and temporal terms constrain the same motion fit. CAHI4D achieves the lowest errors among the compared methods on Hi4D, 3DPW, and CMU toddler. In particular, on CMU, Root MPJPE and Joint-PA MPJPE fall by 21.0% and 49.1%, respectively. Code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Monocular 4D human reconstruction, cross-age human interaction, multi-person mesh recovery, geometric anchoring, contact-aware motion refinement.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_oVHyOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents CAHI4D, an optimization-based framework for monocular 4D reconstruction of cross-age human interactions. The method targets the scale-depth ambiguity caused by age-dependent body morphology and uses validated interpersonal interactions to transfer geometric constraints from a reliably reconstructed participant to an underconstrained one. Its main components include Depth- and Interaction-Guided Geometric Anchoring (DIGA) for relative placement and scale estimation, age-aware shape fitting with Anny-SOMA, and Kinematic-Topology-Guided Interaction Correction (KTIC) for interaction-aware motion refinement. Experiments on Hi4D, 3DPW, and CMU toddler sequences show consistent improvements, with particularly large gains in relative adult-child geometry on the CMU subset.

Critical strengths
  1. The problem setting is an interesting and relatively underexplored problem of cross-age multi-person 4D reconstruction.
  2. The method achieves strong results, particularly on adult-child reconstruction, with useful ablations supporting the main components.
Critical weaknesses
  1. First of all, the high-level presentation of the paper is difficult to follow. Both the abstract and the introduction present numerous technical components before clearly defining the problem decomposition and the overall method logic. In particular, the paper does not clearly explain why the reconstruction problem is decomposed into geometry, shape, and motion stages, which specific failure mode each stage addresses, or why these stages must be performed in this order. As a result, key concepts such as the reference participant, interaction-guided placement, scale correction, shape fitting, and KTIC appear abruptly and are difficult to connect into a coherent pipeline.
  2. The motivation for the core scale-depth ambiguity is not sufficiently developed. Although this ambiguity appears to be the central problem addressed by CAHI4D, the introduction only briefly mentions it and does not clearly show how existing methods fail under cross-age morphology or why interaction-based geometric anchoring is necessary. The subsequent introduction of a “reliable depth reference” therefore feels abrupt rather than naturally derived from the problem analysis. The presentation is further complicated by the introduction of the “Anny–SOMA” representation without clearly stating whether it is an existing representation or a combination constructed in this work, nor why combining Anny and SOMA is necessary for the proposed pipeline.
  3. The method relies on interaction cues to correct inaccurate depth and placement, but these interaction hypotheses are only accepted after checking the current 3D geometry, including surface distance, normals, depth ordering, and temporal consistency. If the initialization is already substantially wrong, I am not convinced that the correct interaction can even pass these checks. In other words, the method may only work when the initial geometry is already reasonably close to the solution.
  4. Once DIGA fixes the geometry, shape fitting cannot revise the placement, and the motion stage only allows very limited root updates. This design may prevent later evidence from correcting mistakes made by earlier stages and could simply lock an incorrect geometric anchor into the final reconstruction.
  5. The interaction event in KTIC updates only one designated owner, while the other participant is detached. This is a strong assumption, especially for adult-adult interactions where both reconstructions may be inaccurate, and there is no obvious reason why only one person should absorb the correction.
  6. The scale correction still depends heavily on the quality of the initialization. DIGA estimates scale by preserving the projected extent before and after depth/placement refinement, and the resulting stature is then enforced during shape fitting. Therefore, if the initial body extent or morphology is already wrong, which is exactly plausible for occluded children, the final scale may simply inherit that error.
Questions for authors
  1. The author should carefully reorganize the high-level presentation of this paper, improving the readability and emphasizing the core challenges and motivation of the proposed method. The current presentation is really hard to follow.
  2. Could the authors perturb the initial depth/scale and report how interaction-admission recall and final reconstruction accuracy change?
  3. Could the authors provide the results of comparing the current sequential optimization with an alternating refinement that allows geometry to be revisited after shape or motion fitting?
  4. How robust is CAHI4D when the initialized body scale or phenotype is inaccurate?

I do think the problem studied is valuable and interesting, and the visualization results are visually pleasing. However, readability remains a major problem, which lowers my rating. Besides, the methodology is somewhat like an engineering composition of several rules that require a bunch of assumptions. These assumptions are not clearly discussed or listed in the paper, which raises my concerns about the method's applicability. I would really like to increase my score if the authors could provide more evidence to support their claim and design.

3 comments

First, I would appreciate it if the authors could provide a revised manuscript with the changes highlighted to facilitate the review.

I appreciate the authors' efforts to reorganize the Abstract and Introduction and clarify the motivation behind individual components. While the roles and motivations of these components are now clearer, I still find the high-level presentation somewhat fragmented. The revised manuscript presents geometry, shape, and motion reconstruction as largely separate challenges, followed by a sequential description of the corresponding modules. As a result, the proposed framework reads more like an engineering-oriented system that combines several task-specific solutions rather than a unified approach driven by a clear technical insight.

My main concern is that the paper does not sufficiently explain how these challenges are fundamentally connected, why they require a unified solution, or what motivates this particular system design. Although Table 6 provides empirical support for the optimization order, it does not fully establish the underlying rationale. Without a clearer connection between the identified challenges and the overall method design, the technical novelty of the framework remains difficult to appreciate.

As a suggestion, I believe the paper could benefit from a more unified narrative that emphasizes the connections and dependencies among geometry, shape, and motion reconstruction rather than presenting them as separate challenges. This would help readers better understand the motivation behind the overall framework and more clearly recognize its technical contributions.

Regarding the technical details, my concerns have been partially addressed. I appreciate the authors' clarifications and acknowledge the substantial improvements over VIMO initialization. However, I remain concerned about the sensitivity of the proposed framework to inaccurate initial geometric estimates, particularly regarding interaction admission and body scale correction. While Table 1 demonstrates overall reconstruction improvements, it does not directly evaluate the robustness of these mechanisms. Similarly, Figs. 7 and 8 mainly illustrate age-dependent body proportions rather than robustness to initialization errors. Since interaction admission depends on intermediate geometric estimates and the recovered stature is fixed during subsequent shape fitting, a sensitivity analysis with different initial depth, scale, or phenotype estimates would provide more convincing evidence of the framework's robustness.

Overall, I appreciate the authors' efforts and the additional clarifications and experimental results. I am willing to slightly increase my rating in recognition of these improvements, although I hope the authors can further address the remaining concerns to strengthen the manuscript.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer oVHy, we hope that our paper revision and the above responses address your concerns. Please let us know your thoughts, and we are more than happy to answer any further questions.

Reviewer_DSGpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The title of the paper is cross-age human interaction 4D reconstruction, but the proposed approach appears to focus on solving a different problem. Looking at the example shown in Fig. 1, the reason that the existing methods may fail is because the person (child) does not interact with the floor, thus, per-person depth cues might be misleading (especially close-interaction scenarios). This is not age specific, as long as one or more persons do not have scene contact (with floors/stairs) together with close interactions, the failures may happen. Thus, there is a mismatch between what the proposed approach addresses (lack of scene contact, close interactions, occlusion) and the open problem the authors claim to address (cross age, age-dependent morphology presents challenges).

The mismatch becomes more obvious in terms of the claimed challenge and the proposed approach. It is stated that age-dependent morphology makes the relative geometry difficult to infer. However, the proposed solution solves different problems. The proposed approach includes using a VLM to provide interpersonal support and contact hypotheses based on visual cues. And the idea of the proposed DIGA module is to use a reliable scene-supported reference and validated interactions to optimize the placement of the ambiguous subject.

Quantitative comparisons on three datasets are provided, where the proposed approach achieves comparable performance to the existing methods on the Hi4D and 3DPW datasets, and improved performance is observed on the CMU dataset (when Anny–SOMA body model is used). To do a fair comparison, the performance of different methods using the same body model is supposed to be compared, in this case, the proposed approach does not show significant improvement when using SMPL as the body model. The performance of using Anny–SOMA body model is reported for the proposed approach, but no performance is provided if using SMPL or SMPL-X with the proposed modules (DIGA, ACSF, etc). Without such information, we cannot rule out the possibility that the performance improvement is largely because of the use of Anny–SOMA body model.

Critical strengths

This submission shows that the authors understand the development in the field well, several recent models are utilized in the proposed approach. For example, the body model used in the proposed approach, Anny–SOMA, is based on a recent ECCV 2026 paper and an arXiv paper released early 2026. In the quantitative comparison, several baselines are papers published earlier this year including ICLR, CVPR, ECCV.

Critical weaknesses

There is a mismatch between the claimed problem this paper focuses on addressing and what the proposed approach actually addresses. It is stated that age-dependent morphology causes the relative geometry to be difficult to infer. However, the proposed approach focuses on using a VLM for interpersonal support and contact hypotheses, and making use of reliable scene-supported reference and validated interactions to adjust the subject with high uncertainty.

The comparisons do not support the proposed approach's effectiveness when the same body model is used. For example, when the same body model is used (SMPL), the proposed approach shows similar performance to the existing methods in Table 1. The improved performance is only observed when a different body model (Anny–SOMA) is used. Qualitative results (Fig. 4) provide a clear illustration to highlight the effectiveness of using a different body model (Anny–SOMA), which shows that existing methods like Human3R, CloseAPP, JOSH typically predict the individuals on a similar scale. As a result, these methods usually have to place the child to be far from the actual position. It is possible that benefiting from the use of Anny, this problem is mitigated. Without a comparison using the same body model, we cannot rule out this.

Questions for authors

Questions:

  • Please clarify and highlight that the proposed approach indeed solves the claimed problem (age-dependent morphology).

  • Discuss the reasons that the performance improvement is limited on Hi4D and 3DPW is relatively small even Anny–SOMA is used as the body model (excluding the Inter. metric).

  • Consider adding a comparison using the same body model to understand how the proposed approach performs compared to the existing methods (all using SMLP/SMPL-X, or all using Anny for example).

Suggestions for future work:

  • The proposed approach is potentially useful to address the challenges presenting in close interactions and occlusion scenarios. Focusing on solving these problems appears to align with the proposed approach.
4 comments
AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer DSGp, we hope that our paper revision and the above responses address your concerns. Please let us know your thoughts, and we are more than happy to answer any further questions.

Thanks for submitting the responses to my questions. The added results of using the same body model provide the comparison when the same body model is used.

Regarding the issue of mismatch between the claimed problem this paper focuses on addressing, the authors revised the paper and added a discussion.

Considering the proposed approach (major components and problems being addressed), focusing on age seems to limit the scope of the paper. For example, the example in Figure 1 is not limited to scenes between child and adults, holding a person (with scale variation) is common in cheerleading or acrobatics, age is just a special case.

Also, this issue is a concern shared by the reviewers. Reviewer 1SrD mentioned in W4 about the limited contribution of age-dependent shape fitting. Reviewer oVHy pointed out how these challenges are fundamentally connected.

That said, a major revision of the paper (title, abstract, and introduction) would be helpful to accurately introduce the motivation and scope, the proposed approach is potentially useful for a broader community.

AuthorsOpenReview ↗

Response to Reviewer DSGp (Follow-up 1)

Dear Reviewer DSGp,

Thank you for your thoughtful feedback and for acknowledging our additional comparison using the same body model. We also appreciate your suggestion to clarify the broader motivation and scope of our work. We agree that the underlying reconstruction challenges are not exclusive to cross-age interactions. More generally, recovering interacting humans from monocular video requires resolving coupled ambiguities in relative placement, body proportions, and motion within a shared 3D space. Cross-age interactions provide a particularly challenging setting because differences in body proportions make these ambiguities more pronounced.

CloseApp (CVPR 2025) and Anny-Fit (CVPR 2026 Findings) illustrate how a particular challenge can motivate a method with broader uses. CloseApp focuses on close interaction, where occlusion makes motion difficult to recover, and also evaluates its motion reconstruction on the interactive sequences of 3DPW. Anny-Fit explicitly addresses the difficulty of fitting children with adult body priors, while its framework recovers adults and children together in multi-person scenes.

Our cross-age emphasis follows a similar rationale. While the underlying reconstruction challenges extend beyond cross-age interactions, our framework includes specific designs to address the age-dependent morphological variations encountered in adult–child interactions. We therefore retain the cross-age emphasis in our title, as it reflects both the motivating challenge and the corresponding methodological design, rather than defining the limits of the framework's applicability. In particular, even after recovering a child's position and stature, we still need to recover the child's proportions to determine where the two bodies can touch as they move. Uniformly scaling an adult template may match the child's stature while placing the contact surfaces incorrectly, causing subsequent pose optimization to compensate for shape errors. To address this, we incorporate age cues to guide morphology fitting at the anchored stature, providing more accurate contact surfaces for subsequent interaction-consistent motion refinement. This age-aware design forms part of our unified treatment of relative placement, morphology, and motion.

Our experiments further support both the age-specific motivation and the broader applicability of the framework. Fig. 7 compares uniform scaling with age-aware fitting on the same CMU toddler frame at matched stature, while Fig. 8 illustrates how head and shoulder proportions vary across age settings even at a fixed height of 1.2 m. On the selected adult–toddler clips, adding the shape-fitting stage after anchoring reduces Joint-PA MPJPE from 83.2 to 79.8 mm (Table 1). Table 4 further measures the smaller incremental benefit of incorporating body-attribute cues into geometry-only fitting, reducing the error from 79.9 to 79.6 mm.

At the same time, our anchoring strategy is not restricted to cross-age interactions. It selects the reference participant based on image, depth, and scene evidence, allowing the same strategy to operate when both participants are adults. Table 1 demonstrates the framework's effectiveness on broader interaction scenarios: relative to initialization, the full framework reduces cross-person joint-distance error from 106.9 to 71.1 mm on Hi4D and from 232.2 to 93.1 mm on 3DPW.

To make this distinction clearer, we will further revise the manuscript, particularly the Abstract and Introduction, to better articulate the general reconstruction challenges, their connections, and why cross-age interactions provide a representative and particularly challenging setting for studying them. We will also clarify the broader applicability of the proposed framework beyond cross-age scenarios, while explaining how its age-aware components specifically address the morphological challenges that motivate our work.

AuthorsOpenReview ↗

Supplementary Response to Reviewer DSGp (Follow-up 1)

Dear Reviewer DSGp,

In our reply to Reviewer 1SrD (W4), we explained what age-aware shape fitting adds once a child’s stature is established. In our reply to Reviewer oVHy (Follow-up 1), we explained how placement, proportions, and pose jointly determine where two bodies can touch as they move.

We believe cross-age interaction deserves attention in a field that has largely focused on adults. Age-related differences in limb and torso proportions affect how one person can reach or support another. CAHI4D was designed around this connection between body shape and interaction, which is why we retain the cross-age emphasis.

Cross-age interaction can also occur between adults of different ages. Taken together, the adult–adult results on Hi4D and 3DPW and the adult–toddler results on CMU provide stronger support for our central claim that a single framework can reconstruct human interactions across ages.

Reviewer_1SrDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CAHI4D, a framework for monocular cross-age human interaction reconstruction. It addresses scale–depth ambiguity and interaction inconsistencies through interaction-guided geometric anchoring (DIGA), age-dependent shape fitting (ACSF), and kinematic-topology-guided motion correction (KTIC). Experiments on Hi4D, 3DPW, and CMU Toddler show improved reconstruction accuracy over existing methods, particularly for adult–child interactions.

Critical strengths
  • S1. The paper addresses an important and challenging problem of reconstructing cross-age human interactions, which remains relatively underexplored in existing work.
  • S2. The proposed method achieves consistent improvements over existing baselines across multiple datasets, with particularly substantial gains on the CMU Toddler dataset.
Critical weaknesses
  • W1. Insufficient Validation of VLM-Based Interaction Reasoning. The reliability of VLM-predicted contact/support hypotheses is not sufficiently evaluated. Furthermore, it remains unclear whether VLM-based interaction reasoning actually improves 3D reconstruction accuracy. Additional analysis of VLM prediction accuracy and comparisons with geometry-based contact estimation without VLM reasoning are needed.
  • W2. Limited Cross-Age Evaluation. Although the paper emphasizes cross-age reconstruction, the evaluation is limited to adult–toddler interactions from a small CMU subset (5 sequences, 13 clips, 2,078 frames). The limited dataset size and lack of diverse age groups and age combinations make it difficult to assess the generalizability of the proposed method.
  • W3. Lack of Interaction-Specific Evaluation Metrics. Despite focusing on human interaction reconstruction, the evaluation primarily relies on pose and joint-distance errors. Although PD-MAE is reported in the ablation study, additional metrics such as contact accuracy, interpenetration, and temporal contact consistency are needed to directly evaluate the geometric and physical plausibility of reconstructed interactions.
  • W4. Limited Contribution of Age-Dependent Shape Fitting. Table 3 shows only marginal improvements (0.2–0.3 mm) over the geometry-only baseline, making the contribution of age-dependent shape fitting unclear. Additional evaluations of body proportions and shape accuracy are needed to demonstrate its effectiveness beyond geometric scaling.
Questions for authors

I would appreciate it if the authors could address the concerns raised in the Weaknesses section, particularly regarding the reliability and effectiveness of VLM-based interaction reasoning, the limited cross-age evaluation, the lack of interaction-specific metrics, and the contribution of age-dependent shape fitting.

4 comments
AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer 1SrD, we hope that our paper revision and the above responses address your concerns. Please let us know your thoughts, and we are more than happy to answer any further questions.

Dear Authors,

Thank you for your detailed responses and additional experiments. I appreciate that the new results have substantially addressed my concerns regarding VLM-guided interaction reasoning (W1) and interaction-specific evaluation (W3).

However, my concerns regarding W2 and W4 remain. The cross-age evaluation is still limited to a small adult–toddler subset, making it difficult to assess generalization across different age groups and age combinations. Additionally, while the authors clarified the contribution of ACSF and demonstrated improvements in joint reconstruction accuracy, direct quantitative evidence of improved age-dependent body proportions or shape accuracy is still lacking.

Since these concerns are closely related to the paper's central cross-age contribution, I will maintain my original rating.

Thank you again for your efforts and clarifications.

AuthorsOpenReview ↗

Reply to Reviewer 1SrD (Follow-up 1)

Dear Reviewer 1SrD,

Thank you for your follow-up and for acknowledging that the additional experiments substantially addressed W1 and W3. We discussed the two remaining points in our earlier replies and briefly clarify what the results establish.

Age coverage (W2). Public datasets of dynamic adult–child interactions with 3D annotations for both participants remain scarce, particularly across multiple child age groups. Cross-age interaction also includes interactions between adults of different ages. The adult–adult results on Hi4D and 3DPW therefore complement the adult–toddler results on CMU and together support the framework’s ability to reconstruct interactions across ages. We agree that performance across additional child age groups and specific age gaps remains to be evaluated.

Age-dependent shape (W4). Table 1 shows that age-aware fitting improves joint reconstruction after placement and stature are anchored. Figs. 7 and 8 illustrate how body proportions change at matched stature. We agree that direct comparison with measured child proportions would provide quantitative evidence of shape accuracy beyond the joint-level results.

Thank you again for your careful assessment.

AuthorsOpenReview ↗

Supplementary Reply to Reviewer 1SrD

Dear Reviewer 1SrD,

We agree that broader age coverage and direct measurements of children’s body proportions would strengthen the evaluation. Such data remain scarce across the field. Capturing interactions across age groups with independent 3D measurements would be a valuable study in its own right. Following Anny-Fit, we used the CMU Panoptic toddler sequences for quantitative evaluation.

We believe cross-age interaction deserves attention because body proportions affect how people reach for and support one another. In the adult–child examples compared in Fig. 4 and our video demo, only CAHI4D consistently recovers a plausible child scale while preserving the observed contact.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.