ManiPose: Manifold-Constrained Multi-Hypothesis 3D Human Pose Estimation

We propose ManiPose, a manifold-constrained multi-hypothesis model for human-pose 2D-to-3D lifting. We provide theoretical and empirical evidence that, due to the depth ambiguity inherent to monocular 3D human pose estimation, traditional regression models suffer from pose-topology consistency issues, which standard evaluation metrics (MPJPE, P-MPJPE and PCK) fail to assess. ManiPose addresses depth ambiguity by proposing multiple candidate 3D poses for each 2D input, each with its estimated plausibility. Unlike previous multi-hypothesis approaches, ManiPose forgoes generative models, greatly facilitating its training and usage. By constraining the outputs to lie on the human pose manifold, ManiPose guarantees the consistency of all hypothetical poses, in contrast to previous works. We showcase the performance of ManiPose on real-world datasets, where it outperforms state-of-the-art models in pose consistency by a large margin while being very competitive on the MPJPE metric.

Paper

References (55)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer fpqy3/10 · confidence 4/52024-07-09

Summary

This paper presents a method to estimate 3D human keypoints from a sequence of monocular 2D keypoints observations. It builds upon an existing sequence-to-sequence architecture (MixSTE), with a different output parameterization exploiting a kinematic skeletton prior, and different training losses. Lengths of the skeletton bones are predicted for the whole sequence to ensure consistency across frames (and maybe also left/right symmetry of the skeletton), and five 3D pose hypotheses with associated scores are predicted for each frame, parameterized as a list of 3D relative orientation for each bone with respect to its parent in the kinematic tree. The authors develop theoretical arguments regarding the benefits of enforcing such structural priors in the predictions, and illustrate with a toy example the interest of having multiple predictions in case of ambiguous multimodal output. They validate their approach on Human3.6M and MPI-INF-3DHP datasets.

Strengths

The motivation for exploiting bone lengths constraints is well expressed, with a clear and detailed discussion provided in Section 4. The discussion of experimental and ablation results is insightful and shows – in a setting dependent on an oracle – benefits of the proposed approach.

Weaknesses

The idea of enforcing body priors (constant bone length here) is not novel and has actually been heavily exploited in a whole line of work relying on more advanced parametric models such as SMPL [100]. This line of work would deserve being considered in the paper, as it encompass approaches suitable for 2D-to-3D sequence lifting such as e.g. [101]. The authors present a pose space consisting in 3D coordinates of joints linked by some rigid segments. Based on this definition, a natural pose parameterization would consist in the 3D direction of each segment, yet the authors chose to overparameterize poses by using relative 3D bone orientation instead. I understand that such choice can have practical benefits in term of biomechanical constraints and additional supervision signal when ground truth data is available, but such choice should be properly motivated, discussed and ablated in the paper. The authors describe two ways of aggregating results L247 but do not state which one they use for MPI-INF-3DHP, and they only report oracle results on Human3.6M and for the ablations. In my understanding, pose hypotheses are selected independently for each frame and there are no temporal terms in the training objectives or aggregation method. Since the proposed approach deals with temporal sequences, it would be worth evaluating the temporal consistency of the predictions, through qualitative video examples and quantitatively e.g. using joint acceleration metrics. Having multiple hypotheses for each frame brings combinatorial questions worth discussing in my opinion. References: - [100] Loper at al., “SMPL: A Skinned Multi-Person Linear Model”, at SIGGRAPH Asia 2015. - [101] Baradel et al., “PoseBERT: A Generic Transformer Module for Temporal 3D Human Modeling”, in TPAMI 2022.

Questions

See the weaknesses section for a list of suggestions.

Rating

3

Confidence

4

Soundness

2

Presentation

3

Contribution

1

Limitations

N/A

Reviewer 5BLY8/10 · confidence 4/52024-07-11

Summary

This paper proposes a MCL-based framework for multi-hypothesis 3D human pose estimation. This framework predicts skeletal parameters so that the predicted 3D poses in a sequence are constrained to one smooth manifold. To prove the superiority of such a framework, the paper presents detailed theoretical analysis on the drawback of unconstrained single-hypothesis HPE and why MPJPE alone is not enough for pose evaluation. The experiments show the proposed framework is capable of keeping the consistency of predicted poses and achieving state-of-the-art MPJPE in the meantime.

Strengths

* Simple and reasonable manifold representation. The proposed framework keeps the predicted human pose on the target manifold by representing the human pose with bone lengths and orientations, and the 3D pose is a direct inference from forward kinematics. The manifold is represented by the kinematics itself. * Inspiring theoretical analysis on basic problems in 3D HPE. The paper arrives at some theoretical conclusions (line178-183), along with detailed proofs. They can provide some refreshing ideas on the innate drawbacks of traditional loss functions and MPJPE metrics. * Good performance under both MPJPE and consistency measures, as validated in Table 2 and 3.

Weaknesses

* Theoretical analysis on the advantage of multi-hypothesis methods over single-hypothesis ones could be added. Specifically, why a **constrained multi-hypothesis** method performs better than an **unconstrained single-hypothesis** method in MPJPE? Though this is already validated by the experiments, I personally believe it would make the paper more solid if the authors could make this analysis. Minor problem: * In Fig.4 (C) and (D), it is not quite clear how the estimations (crosses and triangles) correspond with the inputs (black dots). There might be some unexpected shifts, as the projections of the predicitons do not strictly align with the inputs (like in B).

Questions

What is the quality of the score for each hypothesis? If the multiple hypotheses are fused to one (e.g. by taking the one with the largest confidence or taking the weighted average), then how will the MPJPE, MPSCE, and MPSSE change?

Rating

8

Confidence

4

Soundness

3

Presentation

4

Contribution

4

Limitations

Yes.

Reviewer B5U17/10 · confidence 3/52024-07-12

Summary

This paper presents a new method to estimate 3D human pose from 2D observations (lifting). To ensure the body symmetry and temporal consistency, the authors disentangle human skeleton to two parts: temporally consistency bone scales and temporally variable bone rotations. The authors use fancy formulas to prove that, minimizing MSE loss could not gurantee manifold consistency. The quantitative and qualitative results on Human3.6m and MPI-INF-3DHP datasets show the superiority of the proposed method.

Strengths

1. The evalution results in this paper is quite impressive, especially the newly proposed consistency metric. Figure 1 clearly shows the superiority of the proposed method. 2. The authors try to prove the theoretical optimal of the proposed method, which is worth encouraging.

Weaknesses

I am not an expert in manifold theory, therefore my questions only relate to human pose estimation. 1. How to constrain the rotation space during training? 2. The pose lifting method is quite similar to Anatomy3D (bone length + rotations). Can I view this paper an multi-hypothesis extension of Anatomy3D? Why? 3. Previous paper "POSE-NDF: MODELING HUMAN POSE MANIFOLDS WITH NEURAL DISTANCE FIELDS" is similar to this paper in concepts. SMPL naturally guarantees bone length symmetry, and the learnable parameters (rotations and shape parameters) are similar to this paper in its functionality. It would be better to cite it. 4. Suppose that, there is a virtual dataset, all 2D human joints are rendered (projected) from strictly symmetric 3D joints, then, could learning the lifting function on this virtual dataset using MSE loss guarantee the results all lie on manifold? 5. (An optional question) The ground truth 3D joints of Human3.6M datasets come from the marker tracking on body surface, which naturally could not guarantee skeleton length consistency. Why learning symmetric bones yields better results (both Anatomy3D and the proposed methods)?

Questions

1. The citation style is weird. They are not NeurIPS style, please correct them.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors addressed limitations.

Reviewer tj825/10 · confidence 4/52024-07-13

Summary

This paper propose ManiPose, a manifold-constrained multi-hypothesis model for 3D human pose lifting. The authors provide empirical and experimental evidence to show that joint position regression leads to inconsistent skeleton lengths. And they propose to predict globally consistent pose scale and individual joint rotations per frame (rather than joint positions) to constrain the predictions to the pose manifold. Empirical results demonstrates that the proposed ManiPose framework improves the pose consistency.

Strengths

* The paper provides valuable theoretical analysis to support their arguments and provides intuitive toy examples to illustrate the ambiguity in pose lifting. * The paper conducts extensive experiments on H36M and MPI-INF-3DHP datasets.

Weaknesses

* The paper uses a multi-head design to predict multiple hypotheses. This design loses the flexibility of sampling different numbers of hypotheses and limits the maximum number of hypotheses to a small number. This often results in limited hypothesis diversity. In the experimental section, the authors do not provide numerical of visual measurements of hypothesis diversity. * According to the comparison in Table 4, the manifold constraint proposed in this paper sacrifices MPJPE to improve pose consistency, serving as a trade-off approach between accuracy and consistency. Although the consistency is improved, it lags behind the traditional position regression or manifold regularization in accuracy, and does not bring essential improvement (improve both in accuracy and consistency) compared with these two methods. * Missing comparison with two recent multi-hypothesis methods. [1] GFPose: Learning 3D Human Pose Prior with Gradient Fields. [2] DiffPose: Toward More Reliable 3D Pose Estimation.

Questions

Please review the Weaknesses Section. If the author can address or respond to the above issues well in the rebuttal stage, I will consider increasing my score.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

As the authors discussed in the Limitations Section, they used forward kinematics to obtain joint positions, which can lead to error accumulation.

Reviewer B5U12024-08-10

The authors rebuttal has clarified my concerns carefully. I also notice that the authors add some experiments according to the comments of other reviewers, which makes the evaluation stronger. If the area chairs could guarantee the correctness of mathematical derivation, I think it would be a good choice to accept this paper.

Reviewer tj822024-08-10

Thanks for the authors' response. The response addresses some of my concerns. However, I disagree with the authors' statement that "The diversity in the predicted poses can be measured through Oracle MPJPE performance." MPJPE and diversity reflect two different aspects of the generated results: the accuracy of the best-matching pose and the ability to generate other plausible poses. If we only focus on MPJPE, it aligns more closely with the goal of single-hypothesis pose estimation rather than multi-hypothesis pose estimation.

Reviewer B5U12024-08-10

I think the authors' claim in the rebuttal rooted in the conclusion of "Winner-takes-all learners are geometry-aware conditional density estimators". Maybe the authors could elaborate on this point by showing how to map the conclusion of "winner-takes-all ..." paper with the claim here. From my point of view, the diversity has been demonstrated in Figure.6. Such visualization of diversity has also been used in previous papers like HMFormer. In fact, I think this claim is not totally unreasonable. For example, for a side view person whose left arm is totally occluded, a good predicted distribution (with proper or enough diversity) should cover the pose as close to GT as possible, which would result in a better MPJPE in winner-takes-all evaluation scheme. If it is not overfitting, then the predicted distribution must have a good diversity, at least covering the one close to GT. If the training dataset is large enough, it can not overfit for the occluded parts. Maybe the authors could demonstrate such cases by visualization, which would be much more convincing. I would also like to listen to the advice of other reviewers and the authors about this good question.

Authorsrebuttal2024-08-10

Answer to comments of reviewers tj82 and B5U1

We thank the reviewers for their prompt answers. Reviewer B5U1 is correct in interpreting Figure 6 and in saying that our answer is connected to [A] "Winner-takes-all learners are geometry-aware conditional density estimators". But we understand reviewer tj82's concern. The reason why we say that **oracle MPJPE** is one way of assessing diversity is because it writes as $$\frac{1}{N} \sum_{i} \min_k \ell(f_k(x_i), \mathrm{p}\_i),$$ which **is an approximation when $N$ is large of the quantization error**, also known as distortion: $$\int_{\mathcal{X} \times \mathcal{P}} \min_k \ell(f_k(x), \mathrm{p}) \rho(x, \mathrm{p}) \mathrm{d} x \mathrm{d} \mathrm{p},$$ where $\ell$ is the average joint-wise $L2$ distance in our case. The latter is traditionally used to measure the efficiency of an estimator in summarizing a distribution with few representatives, commonly used to study the K-means estimator for example [C]. So the fact that we achieve better oracle MPJPE than methods requiring a large number of hypotheses (GFPose, D3DP, Wehrbein et al., …) shows that **ManiPose has better quantization properties** than the latter, i.e., it is more efficient in summarizing the diversity of the conditional distribution with fewer representatives/hypotheses. More practically: 1. the extreme case of a model with no diversity at all (e.g., predicting $K$ times the same pose) would lead to an oracle MPJPE = vanilla MPJPE of its single-hypothesis version. This is not what we obtain in our ablation study of Table 4 (1st vs last rows). 2. In the opposite extreme case, a naive way to obtain $K$ very diverse hypotheses would be to use a regular grid of the pose space $\mathcal{P}=\mathbb{R}^{3 \times J}$. The latter learns nothing and is uninformative, but could still achieve better oracle MPJPE if given an unrealistically large number of hypotheses (cf [A] equation 16). 3. As shown in [A], the winner-takes-all learning scheme, used in ManiPose, allows it to sit between these extreme cases by learning an adaptive “grid” made of a few hypotheses, capturing the geometry of the underlying conditional distribution. Results on its quantization optimality can be found in section 5.2 of [A] for example. Of course, oracle MPJPE is not the only way of assessing diversity, which is why **we have provided in our rebuttal additional results** measuring the coverage of ManiPose, which corresponds to the ratio of ground-truth poses whose neighborhood contains at least one generated pose. It is a common metric used in the literature on generative models to analyze diversity. [C] Pages, Gilles, and Jacques Printems. "Optimal quadratic quantization for numerics: the Gaussian case." (2003).

Reviewer tj822024-08-11

Thanks for the authors' response and Reviewer B5U1's comments. I acknowledge that ManiPose demonstrates better quantization properties compared to generative model-based methods like GFPose and D3DP. I suggest that the authors include this comparison and emphasize this point in their revision, as it highlights a unique advantage of their approach. However, I believe that better summarization does not necessarily imply better diversity. I recommend that the authors avoid making such disputed claims in their paper. I also welcome input from other reviewers on this matter.

Authorsrebuttal2024-08-11

We thank the reviewer for taking into account our explanations. We agree that the definition of “diversity” used is indeed important to specify, and will clarify our quantization results in the revised manuscript.

Reviewer tj822024-08-13

Thanks for the authors' response. While I still have concern about the diversity of the results, I recognize that the proposed method offers better summarization ability compared to GFPose and D3DP. Therefore, I am inclined to raise my score. I recommend that the authors include additional metrics related to diversity in their paper and avoid using "Oracle MPJPE" as a proxy for measuring diversity.

Reviewer 5BLY2024-08-12

Thank the authors for addressing my concerns. The theory they prove in the rebuttal is a valuable addition to the contribution of this paper. I have also read the comments from other reviewers and agree that some additional experiments could make this paper more solid. However, I shall vote for acceptance because of the theoretical contributions. If the proofs are guaranteed correct (I only checked the proof sketches, due to limited time and expertise), then the conclusions can be very valuable for the community. Thus, I will keep my rating.

Reviewer fpqy2024-08-13

Thank you for your answer. I acknowledge having read the rebuttal (but I don't have Acrobat Reader unfortunately). For reference, and although this is not directly related to your argument, the continuity theory proposed in [43] is disputed in: Brégier, "Deep Regression on Manifolds: A 3D Rotation Case Study", in 3DV 2021.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC