acceptodds
Under review as a conference paper at ICLR 2027

InfluPose: Learning State-Conditioned Geometric Influence for Category-Level Object Pose Estimation

Abstract

Correspondence-based category-level pose estimation recovers an unseen object’s pose and size from matches between its visible surface and canonical coordinates. Even reliable matches can differ in their geometric role. A broad surface may provide redundant evidence, whereas a small distinctive region can resolve an ambiguity. We introduce InfluPose, which learns this regional influence from interventions on paired complete training geometry. We delete and refill regions, measure the response of a frozen estimator, and combine the resulting changes with a dense geometric proxy. Four target channels describe orthogonal rotation, axial rotation, translation, and size. A point-cloud-only predictor learns these targets from partial observations. It combines influence with correspondence reliability in weighted Kabsch initialization and joint pose-and-size refinement. The pose objective trains the predictor to use this estimator-specific response field for recovery. Complete geometry is required only for training. On unseen object instances from REAL275 and HouseCat6D, InfluPose obtains the highest strict-pose mAP among the compared point-cloud methods. Matched REAL275 ablations show gains from regional supervision over learning the influence coefficients through the endpoint objective alone.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Category-level Object Pose Estimation,Geometric Correspondence,Point Cloud Registration, Robust Geometric Estimation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_kws1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes InfluPose, a category-level pose estimator that complements correspondence reliability with a learned, state-conditioned "geometric influence" field. Concretely, the authors probe a frozen complete-geometry estimator by deleting and refilling M=8 regions, measure the resulting changes in predicted rotation, translation, and size, and blend these finite responses with a dense geometric proxy to form four-channel supervision targets. A partial-cloud predictor is then trained to regress these targets from visible observations, and the predicted influence enters weighted Kabsch initialization and joint pose-and-size refinement, supervised by both direct response losses and the endpoint pose objective. On REAL275 and HouseCat6D, the method achieves the highest strict-pose mAP among compared point-cloud methods.

Critical strengths

++ The authors identify that correspondence reliability does not necessarily imply geometric importance. They design a method to evaluate the influence of each point in the point cloud on the fine-grained components of the predicted pose state, and apply it to point-cloud-based category-level pose estimation, improving performance under stricter evaluation metrics. The motivation is clear, the method design is well aligned with the motivation, and the experimental results validate the effectiveness of the proposed approach.

++ The ablation studies are thorough and well-conducted. In addition, the authors provide the code, which supports reproducibility.

Critical weaknesses

-- While the method performs reasonably well under strict metrics, it falls behind existing methods under more relaxed metrics. What is the reason for this? Could the authors provide analyses on this discrepancy? In addition, could the authors provide an analysis of failure cases?

-- Could the influence-related modules be transferred to other existing methods to improve their performance? This would help demonstrate that the proposed approach is general rather than method-specific.

Questions for authors

See Above

1 comment

Thank you for the positive assessment and for both questions.

W1. Relaxed thresholds and failure cases

Influence only changes how much each predicted correspondence counts when the pose is solved. When the correspondences are about right, better weights give a more precise pose, which gives us the best mAP at 5° on both benchmarks. When they are incorrect, no weighting can fix them, so the instances above 10° remain. Table R1 compares us with HS-Pose (above us at 10°/5cm) on the same detections.

Table R1. Share of instances within each threshold on the same REAL275 detections, ours vs. HS-Pose (%).

BottleBowlCanLaptopMugCamera
5°/2cm, ours84.393.483.179.153.510.0
5°/2cm, HS-Pose63.795.776.355.547.315.7
10°/5cm, ours95.7100.099.894.490.344.1
10°/5cm, HS-Pose97.5100.099.895.587.257.7

At 5°/2cm we lead on bottle, can, laptop and mug. At 10°/5cm these categories are within 2 points of HS-Pose or above it, and only camera is clearly lower.

The failures are of two kinds. The first is camera. The test cameras differ most from the training shapes, and with ground-truth masks our model puts 99.9% of training cameras within 10° but only 44.8% of test cameras. The second is a mug whose handle is hidden. The handle is the only cue for the rotation about the symmetry axis, and most mug errors above 45° occur when less than a quarter of it is visible.

W2. Transfer to other methods

Yes. On DR-Pose, replacing the final pose fit with our influence-based recovery raises the 5°/2cm AP from 41.29 to 44.75, and most of this gain comes from the learned influence (Table R3).

We keep DR-Pose's released weights frozen and add a small influence head on its features, trained with our regional targets and the loss of Equation 5. We then replace its Umeyama fit with influence-weighted Kabsch and one refinement update. No InfluPose network is used at test time.

Table R3. DR-Pose with the influence-based recovery on REAL275 (AP, %, mean of three RANSAC seeds).

Final step of DR-Pose5°/2cm5°/5cm10°/2cm10°/5cmIoU75
Umeyama on RANSAC inliers (original)41.2945.7667.5476.2868.31
Our recovery, all influence set to 142.5746.8767.7876.5268.43
Our recovery, learned influence44.7549.3968.8876.5368.54

Our run of DR-Pose is within 0.41 points of its published results. The gain holds for all three RANSAC seeds and, as in InfluPose, appears at the strict thresholds.

Reviewer_x8qTOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a method that learns regional influence for category-level pose estimation. Some regions look similar under different poses and may occupy a large area of the object. Therefore, the paper tries to identify the important regions: it deletes and refills different regions, measures the response of a frozen estimator, and uses the response to label the geometric influence on the pose. It then trains a predictor on these labels to predict the important regions.

Contributions

The paper identifies an interesting problem. Rather than training or optimizing blindly, it predicts the geometric influence of each region, which is beneficial for the downstream pose estimation. It constructs four-channel targets by combining finite deletion-and-refill responses with a dense geometric proxy, and then couples the predicted influence with correspondence reliability in weighted initialization and joint pose-and-size refinement.

Critical strengths

The problem formulation is interesting. The paper observes that correspondences of similar reliability can play different geometric roles in pose estimation: a broad surface may provide redundant evidence, whereas a small distinctive region can resolve an ambiguity. Instead of training or optimizing blindly, it explicitly predicts the geometric influence of each region, which is beneficial for the downstream pose estimation.

The idea of measuring influence through regional deletion is intuitive. Deleting and refilling regions on complete training geometry and measuring the response of a frozen estimator provides a direct way to label how much each region affects the pose.

Critical weaknesses

Weaknesses

Exposition. Many sentences do not state their purpose and compress several steps of reasoning into one sentence, which makes the exposition hard to follow. For example, "Deleting from a partial cloud measures a response after that sample's own occlusion pattern has already removed points." This sentence states only the premise, while both the purpose (probing the complete training geometry) and the consequence (otherwise, the effect of the deletion cannot be distinguished from the effect of the points already missing due to occlusion) are left implicit.

Notation. Many equations and symbol variants in the method section are introduced without a clear explanation, such as e_y, N_c, R* and t*. Each equation should be followed by a statement that explains it. Some hyperparameters should be moved to the implementation details, since too many symbols are distracting. Moreover, many symbols carry hats, asterisks or subscripts, which seems to exceed the notation needed for this problem. In my opinion, the concept of the method is clear and effective, and its details are not suited to being expressed with so many symbols.

Validation. The experimental settings are unclear. The deletion-based prediction response is the core insight, yet using it alone (Deletion only) performs worse than No learned influence at the 5°/2 cm metric. Shuffled deletion performs slightly better than Deletion only, which contradicts what this paper introduces. The validation section plainly states the results but lacks the necessary explanation and analysis. Therefore, I cannot evaluate which components take effect and which do not.

Questions for authors

Shuffled deletion performs slightly better than Deletion only. How is this consistent with the claim that the deletion responses provide meaningful regional influence?

Clarify the experimental settings of each ablation and provide an analysis of which components take effect.

While it provides an interesting directions, it lacks enough convicing evidence. Provide clear evidence that consistent with method could be better.

Change the method using a clear exposition.

1 comment

Thank you for the careful reading and concrete suggestions.

Q1. Shuffled deletion vs. Deletion only

The two controls are not comparable. The paper did not make this clear, and Section 4.3 of the revision now does. Deletion only gives an example a target only when it is probed, about once in four. Shuffled deletion gives every example a target but puts the responses on wrong regions. The clean comparison is Full vs. Shuffled deletion, which differ only in the region assignment: it is worth 5.60 points. Deletion only is below No learned influence because of its missing targets. With the proxy filling them (β = 1, Table 13), it rises from 51.69 to 55.30.

Q2. Settings and which components take effect

Every component takes effect: removing any of them lowers 5°/2cm AP. Section 4.3 of the revision now defines each control, and Table R4 lists the definitions. Each control keeps Full's data, network, schedule, loss coefficients and frozen estimator, and differs from Full only as listed.

Table R4. Definition of each control in Table 3.

ControlTable in the paperSetting
Full3a, 3bThe target is the measured responses blended with the proxy on probed examples, and the proxy alone on the others.
No learned influence3aThere is no influence field. Every anchor has the same influence, so no target is needed.
No direct supervision3aThe field has no target and learns from the pose losses only.
Proxy only3aThe target is the proxy alone, on every example.
Deletion only3aThe target is the measured responses alone, so an example has one only when it is probed.
Shuffled deletion3aFull's target on every example, with the measured responses assigned to wrong regions.
Trained Kabsch endpoint3bThere is no refinement after Kabsch.
Uniform influence3bThe refinement coefficients are set to one.
Shared influence3bOne coefficient per anchor is shared by all four states.
Full directional operator3bThe full axial/orthogonal operator replaces the diagonal rotation weighting.
Stop-gradient3bPose-loss gradients are blocked through the solver.

Table R5 gives the effect of each component.

Table R5. Effect of each component on 5°/2cm AP (REAL275, Full is 57.73). Losses use unrounded values.

ComponentControl that removes itAP without itLoss
Influence fieldNo learned influence (Table 3a)52.705.02
Direct supervision of the fieldNo direct supervision (Table 3a)54.902.82
Correct region for each responseShuffled deletion (Table 3a)52.135.60
Separate coefficient for each stateShared influence (Table 3b)55.652.07
Any one of the four channelsWithout that channel (Table 5)50.67 to 53.773.96 to 7.06
Training through the solverStop-gradient (Table 3b)21.3236.41

Q3. Evidence for the method

The five results below test the idea. Each cites its table in the paper.

  1. Putting the deletion responses on the correct regions adds 5.60 points (Table 3a, Full vs. Shuffled deletion).
  2. Supervising the field improves accuracy in all three seeds, by 2.70 points on average (Table 4, Full vs. No direct supervision).
  3. Removing any one of the four channels lowers AP by 3.96 to 7.06 points (Table 5).
  4. Deleting the regions ranked highest by influence increases the GT pose error more than random deletion (Table 15, row 1 of Table R6). Deleting the regions with the highest inlier probability is less harmful than random (row 2), and precision-ranked deletion is only slightly more harmful (row 3). Influence therefore finds the regions the pose depends on, and inlier probability does not.
  5. On REAL275 and HouseCat6D, InfluPose has the highest strict-pose mAP among the compared point-cloud methods (Tables 1 and 2).

Table R6. Increase in GT error, top-ranked minus matched random deletion (20% budget, units of Table 15). † 95% interval excludes zero.

Regions ranked byRotationTranslationSize
Influence (Table 15)0.607†0.155†1.068†
Inlier probability-0.273†-0.032†-0.147†
Precision0.1410.041†0.159

Q4. Exposition and notation

We rewrote Section 3 following your comments, without changing the method or results. Your example now states its purpose and consequence. Every symbol is defined at first use, and every equation is followed by a sentence explaining it. Hyperparameter values are moved to the implementation details, and iteration and example indices are removed from Section 3.

Reviewer_bEL4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets category-level pose and size estimation from depth-only partial point clouds, arguing that correspondence reliability doesn't say how much a region matters for each state component. InfluPose freezes a complete-cloud estimator G, deletes 8 FPS regions from paired complete training geometry (refilled by duplication) and measures the change in G's orthogonal/axial rotation, translation and log-size. These responses are blended with an analytic moment-arm proxy and transfered to visible anchors as targets for a 4-channel influence head. At test time influence weights Kabsch and rescales Jacobian columns in a damped robust refinement, trained end to end. Claims: best 5°/2cm mAP among point-cloud methods on REAL275 and HouseCat6D, 144.5 FPS, and ~2.7 pts over endpoint-only training of the same head (3 seeds).

Critical strengths
  1. Clear motivation (Fig. 1); separating reliability from a region's per-state role is reasonable.
  2. Complete geometry and probes are training-only, inference is cheap (A.3.3).
  3. Ablations are more thorough than usual: matched control over 3 seeds (Tab 4), channel removal (Tab 5), sensitivity (Tab 13), stop-gradient, precision/Cauchy (Tab 16).
  4. Object-clustered bootstrap with Bonferroni, and the CI containing zero is reported openly (Tab 6).
Critical weaknesses
  1. Gains are only at strict thresholds. On REAL275, 10°/5cm (81.6) is below five baselines in Tab 1; on HouseCat6D both 10° metrics are 6.2/4.8 pts below ComPose-D. Looks like sharper good cases but more coarse failures. Fixable with an error-distribution/per-category analysis, but it narrows the main claim.
  2. Weak evidence that deletion responses carry the signal (Tab 3a). Proxy only (47.56) and Deletion only (51.69) are both below No learned influence (52.70), so only the blend helps, and Tab 13 shows a sharp peak at the default (beta=0.3: 52.0, M=4/16: 53.2/52.2). Its unclear if Shuffled deletion is blended with the proxy; if so, Full vs Shuffled is the key result and should be stated.
  3. Hyperparameters, ablations and sensitivity seem to be run on REAL275 test, mostly single seed, and Tab 1 reports the best Full seed (57.73) not the mean (57.50).
  4. Sec 4.3 diagnostics (Fig 3, Tab 6, 15) delete regions ranked by predicted influence from the same model whose solver is weighted by that influence, so a larger output change is partly expected by construction. The inlier/precision contrast helps a bit, but Tab 15 has no reliability rankings and an independent estimator would be more convincing.
  5. Per-anchor Jacobian column scaling with unchanged residuals is not a step on any stated objective (A.2.4 admits this), and the exact directional operator does worse than its diagonal approx (53.56 vs 57.73). "Influence" may be more accurately a learned per-state preconditioner.

The score is driven by Weakness 1 (gain only at strict thresholds), W2/W3 (single-source targets hurt, default on a sharp peak picked on test data) and W4. A held-out selection protocol, an explanation of Tab 3a, rankings tested on an independent estimator and an analysis of the 10° drop would raise my score.

Questions for authors
  1. Were M, rho, beta and loss weights selected on a held-out split or on REAL275 test?
  2. Is Shuffled deletion blended with the proxy at beta=0.7? Why do single-source targets underperform no direct supervision?
  3. How many refinement iterations run at test time? A.3.3 mentions one update.
  4. Are the 10° drops driven by gross rotation errors, symmetry handling or specific categories?
  5. In Tab 15, how do inlier- and precision-ranked deletions compare on GT error?
  6. What is the frozen copy in the non-regression loss and when is it updated?
2 comments

We sincerely thank you for the careful and constructive review.

W1/Q4. Gain only at strict thresholds

Influence only changes how much each predicted correspondence counts when the pose is solved. When the correspondences are about right, it makes the pose more precise, which gives us the best mAP at 5° on both benchmarks. It cannot fix incorrect ones, so the instances above 10° remain. This is also the scope the paper claims: the abstract, contributions and conclusion state the highest strict-pose mAP, and Section 4.2 reports the lower 10° metrics.

Table R1. Share of instances within each threshold on the same REAL275 detections, ours vs. HS-Pose (above us at 10°/5cm in Table 1) (%).

BottleBowlCanLaptopMugCamera
5°/2cm, ours84.393.483.179.153.510.0
5°/2cm, HS-Pose63.795.776.355.547.315.7
10°/5cm, ours95.7100.099.894.490.344.1
10°/5cm, HS-Pose97.5100.099.895.587.257.7
  • Specific categories: yes. At 5°/2cm we lead on bottle, can, laptop and mug. At 10°/5cm they are within 2 points of HS-Pose or above it, and only camera is clearly lower. Its test shapes differ most from the training ones.
  • Gross errors: no. Most errors beyond 10° are moderate. Correcting the 1,380 errors between 10° and 45° would raise the mean 10°/5cm AP from 81.56 to 94.10, while correcting the 339 errors above 45° gives only 83.56.
  • Symmetry handling: no. In the symmetric categories (bottle, bowl and can), at least 99.8% of the instances are within 10° in rotation.

W2/Q2. Table 3a

Yes, Shuffled deletion is blended with the proxy at β = 0.70. We added to Section 4.3: "Shuffled deletion keeps Full's blend and probing schedule but assigns the measured responses to the regions in reversed order." This costs 5.60 points in the ablation (Table 3a).

In Table 3a, a single-source target is worse than no target because it is partial or inaccurate. Deletion only gives an example a target only when it is probed, about once in four, and with the proxy filling the rest (β = 1, Table 13) it matches No direct supervision. The Proxy only target agrees only weakly with the measured responses.

W3/Q1. Hyperparameters and seeds

M, ρ, β and the loss weights were not tuned on REAL275. We fixed them before training, M·ρ and β from the design of the probes. With M·ρ = 1 the eight probes cover the shape once, whereas M·ρ = 0.5 leaves half the shape unprobed and M·ρ = 2 removes most points more than once. With β = 0.70 the measured responses rank the regions, and the proxy only orders anchors within a region.

Table 1 uses our default seed, which is the highest of the three, as you note. The other two seeds were run afterwards for Table 4. Over three seeds InfluPose reaches 57.50 ± 0.24, and its lowest seed (57.24) is above every compared method (at most 55.6). Most of the compared methods report only a single seed.

W4/Q5. Reliability rankings, independent estimator

Q5. On GT error, inlier-ranked deletion is less harmful than random deletion, and precision-ranked deletion is only slightly more harmful. Influence-ranked deletion is the most harmful of the three (Table R2).

W4. The effect remains on an independent estimator. For HS-Pose (row 4), it holds in translation and size at all three budgets and in rotation on their average.

Table R2. Increase in GT error, top-ranked minus matched random deletion (20% budget, units of Table 15). † 95% interval excludes zero.

Ranking, estimatorRotationTranslationSize
Influence, InfluPose (Table 15)0.607†0.155†1.068†
Inlier, InfluPose-0.273†-0.032†-0.147†
Precision, InfluPose0.1410.041†0.159
Influence, HS-Pose0.1590.198†0.408†

W5. Influence or preconditioner

We see why the refinement step suggests this reading, and part of it is right: the coefficients do rescale the update. But a per-state preconditioner scales all anchors alike, which changes each update step but not the pose the solver converges to. Our coefficients differ across anchors, so they also change that pose by giving some regions more weight.

We agree that the update does not minimize a fixed objective. This is intended: the solver is trained end to end with the pose loss (A.2.4). The directional form is harder to train. Swapped into the trained model at test time, it lowers AP by less than 1 point, but during training its axial channel receives gradient from a single rotation direction, whereas the diagonal form uses all three axes.

Q3 and Q6

Q3. We use one update at test time (A.3.3) and three in training, as Section 3.3 of the revision states.

Q6. It is the EMA of the model at the end of epoch 30, before the fine-tuning epochs 31 to 50, and is never updated.

The details of both are in our code: https://anonymous.4open.science/r/InfluPose

We add three tables that support our answers to W3/Q1 and Q3 above.

W3/Q1. Hyperparameters

We set M·ρ = 1 and β = 0.70 before training, from how the probes work. Neither was selected on a held-out split or on REAL275. With M·ρ = 1 each point is probed about once, and β = 0.70 balances the two sources of the target (Tables R7 and R8).

A probe measures the effect of the region it removes, and each anchor takes the response of its nearest region. With M·ρ = 1 most points are removed by exactly one probe (Table R7), so each response belongs to one region. With M·ρ = 0.5 half of the points are never removed, and the anchors there take the response of a region that does not contain them. With M·ρ = 2 about three quarters of the points are removed at least twice, so the probes overlap and each response mixes neighbouring regions. The four other probe settings in Table 13, including the M = 4 and M = 16 rows you cite, are these under-covered and over-covered cases.

Table R7. Share of the points of a training shape removed by none, one, or at least two of the M probes (%).

SettingM·ρNeverOnceTwice or more
M = 4, ρ = 0.1250.550.3549.300.35
M = 8, ρ = 0.06250.550.1949.630.19
M = 8, ρ = 0.125 (default)114.6370.9914.37
M = 8, ρ = 0.2520.2124.6275.17
M = 16, ρ = 0.12520.1526.3473.51

The target has two sources, and β sets their mix. The measured responses tell which region matters, but there is one value per region, so they cannot order the anchors inside it. The proxy does order the anchors, but it agrees only weakly with the measured responses on which region matters. With β = 1 all anchors of a region share the same target. With β = 0.30 the proxy takes over, and the region ranking agrees with the measured responses only at about 0.7. With β = 0.70 the agreement stays above 0.9 and the anchors inside each region are still ordered (Table R8).

Table R8. Effect of β on the training target, for the rotation and size channels (the proxy is uniform in the translation channel).

βRegion order agrees with the measured responses (Spearman, 1 = same order)Anchors of the same region get different targets
0.300.68 to 0.73Yes
0.70 (default)0.92 to 0.93Yes
1.001.00No

Q3. Test-time updates

One update is the best choice at test time. It gives the highest strict-pose AP and the highest speed (Table R9).

The number of updates can be changed without retraining, so the table evaluates the same weights with one, two and three updates. With three updates, as in training, the 5°/2cm AP is 57.25, still above every compared method in Table 1 (at most 55.6).

Table R9. Test-time updates on the same weights (AP, %).

Updates at test time5°/2cm5°/5cm10°/2cm10°/5cmFPS
1 (Table 1)57.7362.6975.2781.56144.5
257.2062.4375.0981.69119.9
3 (as in training)57.2562.5575.0881.73102.1

We hope this resolves W3/Q3 and are glad to clarify anything further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.