acceptodds
Desk rejected at ICLR 2027

CDPose: Compositional Diffusion for Whole-Body Pose

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Generating 3D whole-body poses that simultaneously capture body articulation, detailed hand configurations, and expressive facial motion remains a challenging task due to joint rotations evolving in an unconstrained space, where small local errors propagate and escalate across parts. To address these limitations, we propose CDPose, a diffusion-guided whole-body pose model that represents each articulated joint in 6D rotation and converts it inside the generator to an axis–angle before kinematic evaluation. The denoiser is trained under a sub-variance-preserving stochastic differential process with a signal-to-noise–weighted objective. During inference, CDPose produces a single-step Tweedie estimate of the clean pose, with an optional short reverse-time refinement for additional sharpening. The whole-body assembly combines part-specific diffusion heads with a fused residual head that learns cross-part coordination, such as mirrored hand motion and face–torso co-articulation. The coverage and diversity are further improved by mixed training on heterogeneous whole-body and part-only data, using selective masking and supervision. Comprehensive experiments demonstrate that CDPose is robust and versatile across diverse benchmarks spanning body, hand, face, and whole-body pose modeling. In particular, CDPose achieves the lowest errors on the ARCTIC dataset, obtaining a PA-MPJPE of 26.92 mm for body mesh recovery, a PA-MPVPE of 7.10 mm for hands, 2.51 mm for the face, and 23.24 mm for whole-body mesh recovery (PA-MPVPE over all vertices). The code will be publicly released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion prior, Whole body pose, 3D mesh recovery.

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_Zv6hOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes CDPose, a diffusion-based framework for 3D whole-body pose modeling, covering body articulation, hand poses, and facial configurations. The method employs a unified 6D rotation representation to improve optimization stability and converts the predicted rotations into axis-angle representations for compatibility with parametric human models. Experiments across multiple benchmarks evaluate the proposed method on pose generation, denoising, completion, and human mesh recovery, demonstrating improvements over existing pose priors, including VPoser and DPoser-X.

Critical strengths
  1. The proposed framework combines part-specific diffusion priors with a fused residual module to model dependencies among different body regions. Furthermore, the mixed-training strategy allows the model to utilize heterogeneous datasets with incomplete body-part annotations. This design addresses a practical challenge in whole-body pose modeling, where fully annotated training data are often limited, and provides a flexible approach to integrating body, hand, and face information.
Critical weaknesses
  1. Several key components of CDPose, including the 6D rotation representation, diffusion-based pose priors, and Tweedie-based denoising, build upon existing techniques. Although the paper provides ablation studies demonstrating the benefits of individual design choices, it remains unclear to what extent the improvements arise from the proposed compositional architecture rather than the adoption of established representations and training strategies.
Questions for authors
  1. Recent feed-forward human mesh recovery methods, such as SAM3D-Body, have demonstrated strong performance in 3D human reconstruction. While CDPose primarily focuses on learning a generative pose prior, it would be valuable to clarify its practical advantages over or potential complementarity with these approaches. Could the authors discuss whether CDPose can provide additional reconstruction accuracy or robustness when integrated with modern feed-forward models, and how such improvements compare with the additional computational overhead?

  2. Could the authors clarify how facial shape and expression parameters are generated? Specifically, how should the facial generation experiments in Section 4.2 be understood in relation to Appendix D, which states that only joint rotations are diffused?

3 comments

Response to Comment 1 on where the gains come from

We agree, and one correction affects this comment directly. Part of the submitted ablation was invalid, because the rows we read as removing 6D are no-prior baselines, as explained in our response to Reviewer 2. So the submission did not show the benefit of 6D. The matched comparison in Table 6 now does, with a gain of 1.78 mm on ARCTIC hands under our objective.

The composition itself, with frozen part priors, residual fusion, and mixed training, comes from DPoser-X. We change the rotation state and the loss weighting, and Appendix B now separates these as far as the current numbers allow. In whole-body generation, adding the composition with our state and loss fixed lowers FID from 19.23 to 12.81 and raises precision from 0.57 to 0.81, as Table 13 below shows. A comparison with DPoser-X at fixed composition suggests that the state and loss mainly raise precision, but those DPoser-X numbers come from its own evaluator, so we treat that comparison as indicative only.

Excerpt of Table 13, whole-body generation. DPoser-X rows come from its own evaluator

MethodAPD bodyAPD handsFIDPrec.Rec.
DPoser-X base14.452.3490.780.000.00
DPoser-X fused11.781.933.710.320.77
DPoser-X14.082.0413.970.020.81
CDPose-B12.783.4519.230.570.63
CDPose-F14.512.9215.570.660.78
CDPose-M16.465.2412.810.810.92

In fitting, CDPose and DPoser-X share the composition, so the ARCTIC gain comes from the state and loss. On ARCTIC hands, removing the coordination term at fixed state and loss raises the error from 13.47 to 15.55 mm, as shown in Table 7. Switching to axis-angle raises it to 15.25 mm, and switching to DPoser's loss raises it to 15.33 mm, as shown in Table 6. The effects interact, since 6D helps more with our loss, so we report each with the others fixed rather than as shares of one total, as in the table below. All are single runs.

From Tables 6 and 7, one change from the full model, ARCTIC from scratch in mm

Change from the full modelARCTIC AllARCTIC HandsSource
none, full CDPose57.2513.47Tables 6 and 7
coordination weight set to 061.5015.55Table 7
axis-angle state59.2115.25Table 6, row 1
DPoser objective60.1515.33Table 6, row 3
both, DPoser-X recipe in our code61.0416.10Table 6, row 2

Response to Question 1 on feed-forward models such as SAM 3D Body

CDPose is an optimization prior, so using it to refine a stronger initializer is plausible, but we could not test SAM 3D Body within the rebuttal period, so an improvement over it is not established, and the limitations now say so. A strong feed-forward model is the better starting point, since SAM 3D Body reports 5.50 mm PA-MPJPE on FreiHAND against our 9.90 mm from keypoint fitting.

Our existing refinement results, collected in Appendix F.1, include a no-prior control. On ARCTIC from SMPLer-X, refitting without a prior leaves the all-vertex error at 26.33 mm against 26.15 mm for SMPLer-X, while CDPose reaches 23.24 mm. This setting uses ground-truth 2D keypoints and is therefore oracle-assisted. On EHF from CLIFF, the error drops from 56.62 to 48.16 mm, although every prior reaches 49 to 51 mm. So the gain comes from the prior rather than from refitting, and it is largest on the hands.

From Tables 5 and 9, refining a feed-forward output, errors in mm. ARCTIC uses ground-truth 2D keypoints

Benchmark and initializerInitializer aloneRefit, no priorDPoser-X or DPoserCDPose
ARCTIC all vertices, SMPLer-X26.1526.3324.6523.24
ARCTIC hands, SMPLer-X11.2110.347.337.10
EHF body, CLIFF56.62not reported49.0548.16

On cost, each optimization iteration adds one forward pass of the denoiser and no backward pass through it. We removed the earlier runtime table because it mixed our timings with DPoser-X's, so we do not quote a number.

Response to Question 2 on how facial shape and expression are generated

That appendix sentence was wrong and is removed. The face shape prior diffuses 100 FLAME shape coefficients, is trained on MICA registrations, and has no rotations. The face expression prior diffuses the jaw rotation in 6D and 100 expression coefficients and is trained on WCPA. For Sec. 4.2 we sample each prior from noise with the sampler shared by all priors, 1000 Euler-Maruyama steps, and score the samples with the same metrics as the baselines in Table 4. Table 17 now specifies what every prior diffuses, which rotations use 6D, and what is optimized at test time.


Thanks for the detailed response. Despite there are some disagreement about this paper, I still appreciate authors' effort for handling this task. In this day and age, having the patience to study this topic is commendable. I decide to raise my rating to 4.

We sincerely thank the reviewer for their positive response to our rebuttal. We greatly appreciate their encouraging evaluation of our work.

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

The paper proposes CDPose, a diffusion prior for body, hand and face pose. Joint rotations are diffused in 6D and converted to axis-angle inside the model before SMPL/MANO/FLAME decoding. Part-specific priors are trained separately, frozen, and combined by a residual fused head trained on whole-body data; a mixed training scheme masks parts with probability 0.2 and supervises only observed parts. Training uses a sub-VP SDE with a clipped SNR-weighted reconstruction loss; at test time a stop-gradient one-step (Tweedie) estimate with a truncated timestep schedule regularizes optimization-based tasks. Results are reported for pose generation, completion, denoising/IK and mesh recovery on body, hand, face and whole-body benchmarks, with improvements over DPoser/DPoser-X.

Critical strengths

1.The target is worthwhile. A single reusable prior that covers body, hands and face and plugs into generation, completion, inverse kinematics and mesh recovery is useful, and the breadth of tasks evaluated is appropriate for that goal.

2.Some reported gains would be meaningful if they were produced under the stated protocols. For example, ARCTIC whole-body PA-MPVPE from scratch drops from 60.98 (DPoser-X) to 57.25 mm (Table 4), and whole-body completion improves from 21.81/30.99 to 15.61/25.58 mm (Table 12).

Critical weaknesses
  1. The core ablation does not establish the benefit of the 6D representation. In Tables 4, 6, 8, 9 and 10 the "w/o Ingestion" numbers are identical to the "w/o prior" rows of DPoser-X (Lu et al., ICCV 2025), e.g. 72.94/18.80/11.60/86.40 on ARCTIC (DPoser-X Table 8) and 17.71/16.12 on FreiHAND (Table S-9). In the original work these denote optimization without any pose prior, whereas here they are read as the effect of removing the 6D representation (the "44.1%" degradation in Sec. B.4 and the axis-angle bars in Fig. 7), which conflates "whether a prior is used" with "how rotations are parameterized". The ARCTIC axis-angle value of 63.92 mm in Fig. 7(a) matches neither Table 4 nor any experiment defined in the paper, and DPoser's own appendix found 6D slightly worse than axis-angle under identical training (Table S-5 of arXiv 2312.05541v2). A direct axis-angle vs 6D comparison under the same training and test conditions is needed to support the central conclusion.

  2. The contribution beyond DPoser-X is unclear, and the reporting is inconsistent. Part fusion, masked mixed training, single-step regularization, the truncated timestep schedule and the training corpora all overlap heavily with DPoser-X (the paper itself writes "Following Lu et al. (2025)"), and the two new elements, the 6D representation and the Eq. 9 loss weighting, have not been isolated. The main text misdescribes DPoser as modeling 3D joint locations (it diffuses axis-angle rotations); the baseline columns of Table 7 appear shifted by one position relative to DPoser Table 2 (108.57 is "no fitting", not GMM, and the GAN-S values are missing); Table 11 omits DPoser-X-fused, whose FID (3.71) is lower than CDPose's (12.81); and the face experiments give contradictory accounts of the diffused variables and training data (Sec. 4.2 vs App. C.3 and D). Smaller inconsistencies (optimizer, hand data sources, swapped metric definitions, Table 13's mixed-provenance runtimes, an unresolved "DECA ?" citation) add to the difficulty. Together these make the novelty and the experimental conclusions of the current version hard to assess reliably.

Questions for authors
  1. For every table, state precisely what "w/o Ingestion" denotes and how the numbers were obtained, and reconcile them with the identical "w/o prior" rows of DPoser-X. A representation-only ablation (same architecture, data, training budget and test-time schedule; axis-angle vs 6D diffusion) on ARCTIC, FreiHAND and ReInterHand, with a separate no-prior row, would be very helpful, as would an explanation of the 63.92 mm value in Fig. 7(a).

  2. Novelty and provenance relative to DPoser-X. Which modules differ from DPoser-X-mixed? Please give a component table (inherited / modified / new) and an ablation that adds the 6D parameterization and the Eq. 9 weighting one at a time on top of a DPoser-X reproduction, including the sampler and step count used for generation metrics (this would also confirm the sampler behind the APD of 19.03). Please state the provenance (own run vs copied, and hardware) of the numbers in Tables 1–13, check Table 7 against its source, and consider including DPoser-X-fused in Table 11. Please also state which version of DPoser Sec. 2.1 compares against and correct its characterization, since the only Lu et al. entry in the bibliography is DPoser-X, whose Sec. 2.2 models joint rotations and whose Sec. 2.5 already contains part fusion and masked mixed training.

  3. Face and expressiveness claims. What exactly is diffused in the face shape and expression experiments of Table 3 (variables, dimensions, network, training data), given Appendix D? Since ARCTIC evaluation sets facial expressions to zero (App. C.4), what supports the claim of expressive face modeling? Please also label the abstract's headline numbers with their initialization (SMPLer-X) and keypoint source (ground-truth 2D).

4 comments

Response to Comment 1 on the 6D ablation

The reviewer is right. The rows labelled without ingestion are DPoser-X's published results without any prior, not an axis-angle version of CDPose, so they say nothing about the 6D state. We withdraw the 44.1 percent attribution, which Sec. C.3 now describes as the value of having a prior, and we removed the 63.92 mm value, which does not come from a documented experiment. The axis-angle bars in Fig. 7 now come from the matched runs below. Every table marks each baseline as quoted or reproduced, as listed in Table 15.

We ran the matched experiment the reviewer asks for, a two by two design that crosses the rotation state, axis-angle or 6D, with the objective, DPoser-X's or ours, while keeping data, fusion, masking, training budget, initialization, keypoints, and fitting fixed, with no-prior fitting reported separately in Table 6 below. With our objective, 6D lowers every error, from 10.15 to 9.90 mm on FreiHAND, from 1.40 to 1.22 mm in ReInterHand inverse kinematics, and from 15.25 to 13.47 mm on ARCTIC hands. With DPoser's objective 6D also helps but less, from 16.10 to 15.33 mm on ARCTIC hands. DPoser's appendix found 6D slightly worse in its own setting, so we state our result for our setting only. These are single runs, so we report them without confidence intervals.

Table 6 of the revised paper, rotation state crossed with objective, errors in mm

RowVariantFreiHAND det. PA-MPJPEReInterHand IK MPVPE at 2 mmARCTIC AllARCTIC Hands
06D, our objective, full CDPose9.901.2257.2513.47
1axis-angle, our objective10.151.4059.2115.25
2axis-angle, DPoser objective, DPoser-X recipe10.901.7961.0416.10
36D, DPoser objective10.601.5060.1515.33
no prior, from DPoser-X17.713.9572.9418.80

Response to Comment 2 on the contribution beyond DPoser-X and the reporting

We agree that most of the recipe is DPoser-X's, and Table 1 now marks each component as inherited, modified, or new, as shown in our answer to Question 2. We change two things, the rotation state and the training loss. Because the error of the clean-pose estimate is the noise error scaled by the noise-to-signal ratio, our loss in Eq. 9 is the same noise-prediction loss with a different timestep weighting, close to Min-SNR weighting, and the paper now says so. Table 6 isolates the state and the weighting, starting from DPoser-X reproduced in our code, which gives 61.04 and 16.10 mm on ARCTIC over all vertices and on the hands, against DPoser-X's published 60.98 and 15.60 mm. Our weighting helps with either state, for example from 16.10 to 15.25 mm on ARCTIC hands with axis-angle.

The reporting errors are real and fixed. DPoser is described as a prior over axis-angle rotations. Table 13 includes DPoser-X fused, which has the lowest FID, 3.71 against our 12.81. The EHF table, Table 9, now follows DPoser's table exactly, as below, with 108.57 labelled as no fitting, GAN-S at 57.26 and 49.58, and the unsourced NRDF column removed. Our result stays 55.08 and 48.16 mm against DPoser's 56.05 and 49.05 mm.

Table 9 of the revised paper, EHF body PA-MPJPE in mm, baselines in DPoser's column order

InitializationNo fittingGMMVPoserPose-NDFGAN-SDPoserCDPose
From scratch108.5758.3258.0857.8757.2656.0555.08
CLIFF56.6251.0249.3949.5049.5849.0548.16

The face has one specification, in Table 17. The shape prior diffuses 100 FLAME shape coefficients from MICA, with no rotations. The pose prior diffuses the jaw in 6D and 100 expression coefficients from WCPA. The whole-body prior includes the expression but not the shape.

The optimizer, the hand-data list, the swapped metric definitions, and the DECA citation are corrected. The runtime table mixed our timings with DPoser-X's, so we removed it rather than report numbers of mixed provenance. Table 15 gives the source of every number.

Response to Question 1 on the rows without ingestion and the 63.92 mm value

In every table it meant optimization without any prior. In ARCTIC, hand inverse kinematics, FreiHAND, NoW, and face inverse kinematics, Tables 5, 8, 10, 11, and 12, the rows are quoted from DPoser-X and now cite their source tables. The motion denoising table had a row that matched no source. That table is body-only and outside our focus, so we removed it from the revision, and Fig. 7d keeps only our own denoising runs with DPoser's published values as reference. The runtime table is removed. The representation-only ablation is the two by two design in our response to Comment 1, with the no-prior result as a separate row in Table 6 and a separate bar in Fig. 7. The 63.92 mm value could not be traced to a documented experiment, so it is removed, and the matched axis-angle run gives 59.21 mm on ARCTIC in its place.

Response to Question 2 on the modules and the provenance of the numbers

Table 1 now marks each module as inherited, modified, or new, as reproduced below. Inherited are the part priors, the process and schedule, the one-step regularizer, the composition, the mixed data and masking, and the sampler. Modified are the rotation state and the loss weighting. New are the in-generator conversion, the optional multi-step refinement, the sparse projector, and the test-time weight, which we added in this revision. Table 6 adds 6D and the Eq. 9 weighting one at a time to a DPoser-X reproduction, as shown in our response to Comment 1.

Table 1 of the revised paper, components shared with and changed from DPoser-X

ComponentDPoser-XCDPoseStatus
Rotation stateaxis-angle, z-scored6D, all joints incl. jawmodified
Non-rotational state100 expression coefficients, separate shape priorsameinherited
In-generator conversionnoneblock-wise 6D to axis-anglenew
Training lossnoise prediction, sigma squared weightTweedie reconstruction, clipped SNR, Eq. 9modified
Test-time priorone-step, stop-gradone-step Tweedie, stop-grad, optional N-stepinherited, N-step new
Process, schedule, part priorssub-VP, truncated 0.05 to 0.15, frozensameinherited
Compositionpart scores plus residual from part featuressame, Eq. 13inherited
Mixed data, maskingp 0.2, weights 1, 0.2, 0.2, 1, 0.5same weights, plus BEATinherited
Generation samplerEuler-Maruyama, 1000 steps, no correctorsameinherited
Coordination at test timefixedtunable weight w, Eq. 15new in revision
Keypoint mappingdense regressorssparse projector, Eq. 12new
Parameters, part and fused8.28M and none7.2M and 4.1M

Generation uses DPoser-X's sampler, 1000 Euler-Maruyama steps without a corrector, as stated in Appendix E. Our APD of 19.03 is identical to that of DPoser-X's 10-step DDIM variant, so we withdraw the diversity claim built on it, and Table 16 marks the value as under recheck with our sampler.

Table 15 states, for every table, which rows are quoted and from where, and which are our runs. We re-ran no baseline except BUDDI, which we sampled ourselves. This check also showed that the inverse kinematics tables gave baselines at one noise level only. They now show all levels, and DPoser-X's face prior is slightly better than ours at 2 and 5 mm. We now cite DPoser, arXiv version 2, separately from DPoser-X, describe it as a rotation prior, and credit part fusion and masked mixed training to DPoser-X.

Response to Question 3 on the face priors and expressive modeling

Table 17 now specifies every prior, as reproduced below. The face shape prior diffuses 100 Euclidean FLAME shape coefficients and is trained on MICA registrations. The face pose prior diffuses the jaw rotation in 6D and 100 Euclidean expression coefficients and is trained on WCPA. Both use the fully connected backbone of the other part priors. Shape and expression coefficients are never represented in 6D.

Table 17 of the revised paper, what each prior diffuses

PriorDiffused variablesIn 6DEuclideanTraining data
Body21 body joint rotationsallnoneAMASS
Hand15 joint rotations of one hand, left mirroredallnonehand datasets of App. D
Face posejaw rotation, 100 expression coefficientsjawexpressionWCPA
Face shape100 FLAME shape coefficientsnoneshapeMICA registrations
Whole body52 joint rotations, 100 expression coefficientsall rotationsexpressionBEAT2, GRAB, ARCTIC, EgoBody, part-only data

ARCTIC fixes expressions to neutral, so it does not test expression recovery, and NoW tests shape only. Our expression evidence is face-level only, from expression generation in Table 4 and face inverse kinematics on WCPA in Table 12, where we are better than DPoser-X at 1 mm noise and slightly worse at 2 and 5 mm. We therefore no longer claim expressive whole-body modelling, and the limitations say that it remains untested. The abstract now states the SMPLer-X initialization, ground-truth 2D keypoints, and neutral expression behind its headline numbers.


Thank you again for your thorough review and constructive feedback. We hope our responses have adequately addressed your concerns, and we would be happy to answer any further questions or provide additional clarification if needed.

Reviewer_NqNoOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents CDPose, a diffusion prior for whole-body pose that covers body, hands and face together. All joints are kept in 6D rotation and converted to axis-angle inside the generator, right before the kinematic evaluation. The denoiser is trained with a sub-VP SDE and an SNR-weighted objective. At inference it uses a one-step Tweedie estimate of the clean pose, with an optional short reverse-time refinement. On the architecture side there are part-specific diffusion heads plus a fused residual head that learns cross-part coordination, and the model is trained on mixed whole-body and part-only data with selective masking. Experiments cover body/hand/face generation and whole-body mesh recovery on ARCTIC.

Critical strengths
  1. The execution is competent and the scope is broad. One single prior serves body, hand and face generation as well as whole-body mesh recovery, and on ARCTIC it is the best on every region in both the from-scratch and the SMPLer-X-initialized settings (for example All 23.24 vs 24.65 for DPoser-X).
  2. The ablations are organized well too: the 6D-vs-axis-angle comparison is convincing (ARCTIC 63.92 down to 57.25, hands 17.43 down to 13.47), the timestep policy degrades smoothly as the input gets noisier, and the part, fused and mixed training stages each add something.
Critical weaknesses
  1. My main concern is that there is no new idea here. 6D rotation, diffusion pose priors (BUDDI, DPoser), a fusion head and mixed training are all existing pieces, and the three differences the paper lists against BUDDI and DPoser are engineering choices rather than a new principle. The contribution is basically a recombination.
  2. The claims outrun the numbers. DPoser has a better FID on both body generation (0.07 vs 0.13) and hand generation (0.01 vs 0.02), the ARCTIC margins over DPoser-X are modest on the body metric (70.51 vs 73.00, and 26.92 vs 27.87 with SMPLer-X init), and the text calls its APD the highest while its own Table 1 gives NRDF 22.82 against 19.03. So "consistent improvements" and "new state of the art" are stronger than what the tables show.
  3. The evaluation is incomplete for the paper's own positioning. BUDDI is named as a closest work but appears in no table, and the hand and face heads are trained on HaMeR/MICA pseudo-ground-truth whose ceiling is never discussed. Taken together this puts the paper below the bar for me.
Questions for authors
  1. Can you add BUDDI to the comparison on the same tasks? You call it one of your closest works, so leaving it out makes the evaluation incomplete.
  2. How much does training on HaMeR/MICA pseudo-labels limit the hand and face prior? Reporting the upstream models' own error, or a small study with real annotations, would help clarify the ceiling.
  3. On body and hand generation DPoser beats you on FID. What does CDPose offer that DPoser does not, and can you show a downstream case where that difference actually matters?
3 comments

Response to Comment 1 on novelty

We agree that each ingredient exists, and most of our recipe is DPoser-X's. Table 1 marks every component as inherited, modified, or new. We change two things, the rotation state and the timestep weighting of the training loss in Eq. 9, and Table 6 tests each one against a DPoser-X reproduction. On ARCTIC hands, DPoser-X's recipe retrained in our code gives 16.10 mm, adding only 6D gives 15.33 mm, adding only our weighting gives 15.25 mm, and adding both gives 13.47 mm, so each change helps and the two reinforce each other.

We also read the fused prior as independent part priors plus a learned residual correction, as described in Sec. 3.6. This is an interpretation, not a new model, and we no longer present it as a principle. It is useful because the correction can be measured and scaled at test time with one weight, as Table 7 below shows. As the weight grows from 0 to 2, hand symmetry and the jaw and head correlation move toward the real data, while FID, hand diversity, and ARCTIC error are best at the trained weight of 1. A plain fusion head, DPoser-X's included, can be scaled the same way, so the weight is not unique to our model, and we only measure its effect on CDPose. The weight was added during this revision.

Table 7 of the revised paper, coordination weight at test time, no retraining

WeightHand symmetry W1Jaw-head corr. gapFIDAPD handsARCTIC AllARCTIC Hands
0, equals CDPose-B0.690.3119.233.4561.5015.55
1, trained CDPose-M0.360.1612.815.2457.2513.47
20.270.1113.014.6358.4214.09

BUDDI models two interacting people, so it is not our closest work, and the related work now says so. Our direct baseline is DPoser-X, and on ARCTIC CDPose has the lowest error of the compared priors on every region, as shown in Table 5. The introduction is rewritten and the list of three differences is removed.

Response to Comment 2 on claims and numbers

We agree, and the claims now match the tables. Consistent improvements and state of the art are removed, and every result is stated per metric and per region.

In generation DPoser has the lower FID for body and hand poses, and the paper says so first. Our advantage is higher generative precision and nearest-neighbour distance, and for hands higher diversity. The APD sentence was wrong, since NRDF has the highest APD, although with low generative precision under this protocol. We also withdraw the claim that our body samples are the most diverse among high-precision priors, and our body APD is under recheck, as explained in our answer to Question 2 of Reviewer 2.

On ARCTIC the margin over DPoser-X depends on the region, as the excerpt of Table 5 below shows. Fitting from scratch it is 3.4 percent on the body and 13.7 percent on the hands, and with SMPLer-X initialization and ground-truth 2D keypoints it lies between 3 and 8 percent. We now call the body gain modest. Table 16 lists the revised claims and their evidence.

Excerpt of Table 5, ARCTIC whole-body fitting in mm, PA-MPVPE for all, hands and face, PA-MPJPE for body

SettingMethodAllHandsFaceBody
From scratchDPoser-X60.9815.609.7573.00
From scratchCDPose57.2513.479.1270.51
SMPLer-X init.DPoser-X24.657.332.7327.87
SMPLer-X init.CDPose23.247.102.5126.92

Response to Comment 3 on BUDDI and the training labels

Calling BUDDI our closest work was a mistake, and the related work is corrected. BUDDI is a two-person prior without hand or face parameters, so it does not apply to our hand, face, or whole-body tasks. Its body poses can be compared personwise in body generation, and Table 2 now includes it, scored with the same code as every other prior. BUDDI reaches an FID of 0.81 and a precision of 0.46, against 0.13 and 0.77 for CDPose. It is trained on interaction data rather than AMASS, which the paper notes.

Excerpt of Table 2, body pose generation. BUDDI is our run, and the CDPose APD is under recheck

MethodAPDFIDPrec.Rec.dNN
AMASS, real data15.440.0050.860.900.001
NRDF22.820.640.030.996.69
DPoser14.280.070.720.802.63
BUDDI15.260.810.460.544.40
CDPose19.030.130.770.822.58

The term pseudo-ground-truth confused where we load the data with how the labels were made. None of our labels are HaMeR or MICA predictions. Most hand labels are optimization fits to multi-view captures, but those of InterHand2.6M come from NeuralAnnot, a learned annotator. The face labels are FLAME registrations to scans and the FLAME parameters of WCPA. Table 18 gives the source of each label. All are fits, so their fitting error and capture coverage remain a limitation, which Sec. 4 now discusses. ARCTIC is evaluated on held-out sequences whose labels are also fits to motion capture, but the part priors never see them.

Response to Question 1 on adding BUDDI

BUDDI's native task is two-person modelling, which differs from our single-person evaluation. A personwise body-pose comparison is possible in generation, and Table 2 now reports it with the same evaluator as every other prior, as shown in our response to Comment 3. It gives no comparison for hands or expressions. An adaptation to single-person fitting is conceivable, but it would test the adaptation as much as the prior, so we do not report one.

Response to Question 2 on the limits of the training labels

Training on dataset fits rather than network predictions removes the concern about distilling a feed-forward estimator, but fitting error and model bias remain, and Table 18 lists how each label was made. A small study on real annotations would measure this directly, and we have not completed one. The closest evidence we have is ARCTIC, where the hand error drops from 15.60 to 13.47 mm against DPoser-X on held-out sequences that the hand prior never saw, as shown in Table 5. We removed NoW from this evidence, because the values in our NoW table are about ten times those reported by MICA, and we are checking their units and protocol.

Response to Question 3 on FID and downstream benefit

FID compares sample statistics with the training set, while precision and recall measure how many samples lie near the data and how much of the data they cover. For fitting, precision and coverage matter more than matching the mean and covariance. The difference shows when the evidence is ambiguous or noisy. In hand completion from body and face context, CDPose has the lowest best of 10 and mean errors on ARCTIC and BEAT2, below both DPoser-X and DPoser-X fused, as Table 14 below shows. With detected keypoints on FreiHAND the error drops from 10.71 to 9.90 mm, as shown in Table 10, and in hand inverse kinematics at 2 mm noise from 1.71 to 1.22 mm, as shown in Table 8. With clean evidence the two priors are tied, and with ground-truth keypoints on FreiHAND both reach 8.68 mm.

Table 14 of the revised paper, completing a masked hand, MPVPE in mm over 10 hypotheses

MethodARCTIC minARCTIC meanARCTIC stdBEAT2 minBEAT2 meanBEAT2 std
VPoser-X37.3443.244.6027.4935.465.06
DPoser-X21.8130.996.1015.9225.896.04
DPoser-X fused21.5130.375.9615.5124.586.25
CDPose15.6125.586.1712.9421.675.39

We hope these corrections address the concerns behind the reviewer's rating.

Thank you again for your thorough review and constructive feedback. We hope our responses have adequately addressed your concerns, and we would be happy to answer any further questions or provide additional clarification if needed.

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

Related papers

Discussion (0)

Sign in to comment.