acceptodds
Under review as a conference paper at ICLR 2027

MAMHOI: Factorizing Scene-Aware Human-Object Interaction through Affordances

Abstract

Generating realistic human-object interactions (HOI) in complex 3D scenes requires two complementary capabilities: reasoning about interaction feasibility in the environment and synthesizing realistic human-object motion. However, supervision for these capabilities is rarely available jointly at scale. Human-scene datasets provide rich information about environment-aware motion, while human-object datasets capture detailed interaction dynamics, yet paired human-object-scene data remain scarce. We present MAMHOI, an affordance-mediated factorization for scene-aware human-object interaction generation. MAMHOI factorizes scene-aware HOI generation through an explicit motion-affordance interface between scene understanding and motion synthesis: a scene-conditioned model first predicts where and how an interaction can be feasibly executed, and an affordance-conditioned HOI model then generates the corresponding human-object motion. This factorization allows scene understanding and interaction dynamics to be learned from complementary sources of supervision without requiring paired human-object-scene data. Experiments in complex indoor environments show that MAMHOI reduces object–scene penetration while better preserving human–object interaction quality, yielding more realistic and physically feasible scene-aware interactions.

Primary area:
generative models
Keywords:
Human-Object Interaction, 3D Human Motion Generation, Scene-Aware Motion Synthesis

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

This paper proposes MAMHOI, a factorized framework for scene-aware human–object interaction generation. It introduces motion affordance as an intermediate representation to connect scene understanding and HOI motion synthesis, allowing the two components to be trained independently on human–scene and human–object datasets. Experiments demonstrate improved scene compatibility and interaction quality compared with existing methods.

Critical strengths
  • Meaningful problem. Learning scene-aware HOI from separate HSI and HOI datasets addresses the scarcity of paired training data.
  • Intuitive factorization. Motion affordance provides a reasonable interface between scene understanding and interaction generation.
  • Strong quantitative results. The method improves penetration, motion smoothness, and contact metrics on the constructed scene-aware benchmark.
  • Additional validation. The paper includes ablations, an auxiliary evaluation on TRUMANS, and a perceptual user study.
Critical weaknesses
  • Insufficient validation of the learned affordance representation. The paper lacks comparisons against simpler alternatives, such as direct scene occupancy conditioning or path-based affordance construction. It is unclear whether the improvements come from the learned affordance representation or simply from introducing scene-aware conditioning.
  • Definition of affordance. There is a domain gap between the affordance used in training (oracle) and inference (prediction). There is no quantitative evaluation of this gap.
  • Loss of 3D geometric information. Appendix D.3 describes vertically pooling the local occupancy voxels and broadcasting the resulting 2D representation along the height dimension. This discards height-dependent geometric information and may limit the ability to handle complex 3D obstacles.
  • Dependence on external path planning. The method requires a planned object path, and the evaluation uses manually selected feasible start–goal configurations with A* planning. Its robustness to imperfect or physically infeasible paths remains unclear.
Questions for authors
  • Height collapse. Appendix D.3 min-pools each local volume along the vertical axis and broadcasts it back along height, so the condition appears to carry no height information. How does the model then distinguish passing under an obstacle from routing around it?
  • Oracle vs. predicted affordance. Training uses an affordance built from the observed motion, inference uses the predicted one (Section 3.3.2). Could you report Table 2 with the GT-derived affordance as an oracle, to quantify this gap?
  • Planner contribution. What happens with the path removed (text + scene only)?
  • Affordance semantics. The affordance is an AABB around LINGO human-only trajectories (Section 4.1) but is defined as the joint human–object support (Section 3.3.1). How does it capture the wider clearance large objects need?
  • User study scope. All four pairings in Appendix F are MAMHOI vs. CHOIS variants. Could you add InfBaGel, and report significance for the 884 votes?
  • Inference cost. Appendix D.5 specifies 1,000 denoising steps over autoregressive windows. Could you report wall-clock time and memory, given the characterization of HOSIG as expensive in Section E.3?
2 comments

Thank you for highlighting our strong quantitative results and for your helpful suggestions. We address your comments and questions below.

W1: Learned affordance versus simpler conditions

Without affordance conditioning during training, the generator reduces to the CHOIS baseline, which is already included in our comparisons. To isolate the contribution of learned affordance within our trained model, we remove it or replace it with an A* corridor or raw scene occupancy at inference, keeping the generator and planned paths fixed. These variants yield higher scene penetration despite similar path errors, supporting the benefit of learned affordance beyond path control (Appendix E.2, Table 4).

Object-scene penetration depth is averaged per frame.

Spatial conditionHuman-scene penetration rate (%) ↓Object-scene penetration depth (×10⁻³) ↓Object-path error ↓
A* corridor25.9661.7940.1490
No scene condition30.2342.2130.1527
Original scene occupancy18.7251.7860.1500
Ours14.9901.4400.1492

W2 & Q2: Affordance transfer and the oracle gap

At inference, affordance is predicted from the distribution learned on LINGO, while the generator is trained with affordances derived from OMOMO motion. Our factorization transfers coarse spatial support without requiring identical affordances across domains. As a soft condition, it guides motion without requiring strict adherence to its boundaries. The generator combines it with object geometry, the planned path, and the learned HOI prior. The distribution gap remains a limitation, and we will explore better alignment in future work.

The ablation in W1 (Appendix E.2, Table 4) supports the usefulness of this transfer, although it does not quantify the oracle gap.

A matched oracle evaluation for Table 2 is unavailable because the test scenes lack ground-truth HOI sequences corresponding to the planned paths. GT-derived affordances would also expose target-motion information, introducing ground-truth leakage and preventing a like-for-like comparison with standard inference. We have added this clarification in Appendix E.2 (lines 1104 to 1109).

W3 & Q1: Height-dependent scene constraints

We agree that 2D affordances lose height-dependent information, limiting distinctions such as passing under an obstacle versus routing around it. This is a limitation of the current representation, not a fundamental limitation of our factorization. We have clarified this in lines 508 to 516 and will explore richer height-aware representations in future work.

W4 & Q3: Dependence on external path planning

Our formulation assumes a planner-provided object path and does not perform high-level route planning. Routes incompatible with object geometry should be rejected or adjusted by an object-aware planner, as our generator is not guaranteed to correct them. Our evaluation therefore uses feasible start and goal configurations. We clarify this assumption and limitation in lines 512 to 513.

Both factors are trained with the path as an input, so removing it only at inference would test a missing required condition. A meaningful comparison would require training for generation without a prescribed path and evaluating task completion rather than path adherence. We leave joint route planning and interaction synthesis to future work.

Q4: Affordance semantics and object clearance

We have revised Section 3.3.1 to clarify that Factor I predicts coarse human-centric spatial support relative to the planned object path. Because it learns from LINGO without object geometry, it cannot explicitly capture object-specific clearance requirements. Factor II combines this soft condition with object geometry, the planned path, and the learned HOI prior to generate coordinated motion.

Our new per-object evaluation (Appendix E.2, Table 5) compares generation with and without predicted affordance guidance. Across all six objects, this guidance reduces human-scene penetration rates and object-scene penetration depths while maintaining similar path errors.

Entries show Without → With (relative change). Affordance is removed only at inference, with the generator and paths fixed. These results provide empirical evidence that HSI-learned affordances offer useful guidance for HOI generation across different objects, although they do not establish object-specific clearance.

ObjectHuman-scene penetration rate (%) ↓Object-scene penetration depth (×10⁻³) ↓Object-path error ↓
clothesstand31.084 → 12.200 (-60.75%)3.590 → 1.193 (-66.77%)0.1527 → 0.1482 (-2.95%)
floorlamp29.291 → 13.279 (-54.67%)1.651 → 1.535 (-7.03%)0.1553 → 0.1511 (-2.70%)
largebox28.641 → 14.312 (-50.03%)1.666 → 0.944 (-43.34%)0.1530 → 0.1482 (-3.14%)
largetable32.180 → 16.164 (-49.77%)3.425 → 3.152 (-7.97%)0.1487 → 0.1471 (-1.08%)
smallbox26.518 → 15.657 (-40.96%)1.432 → 0.511 (-64.32%)0.1562 → 0.1544 (-1.15%)
whitechair33.691 → 18.327 (-45.60%)1.512 → 1.303 (-13.82%)0.1505 → 0.1463 (-2.79%)

We have clarified this limitation in lines 508 to 516 and will explore object-aware affordances in future work.

Q5: Expanded user study and statistical significance

We have expanded the user study to include InfBaGel and report all six pairings in Appendix F, Table 9. All comparisons are significant at p < 0.001 using two-sided exact binomial tests against a preference probability of 0.5, under an independent-judgment assumption. The test is described in the caption (lines 1245 to 1249), and the main-text results are updated in lines 494 to 507.

Votes and preference percentages refer to Method A. SDF denotes test-time SDF guidance.

Method AMethod BA votes / totalA preferenceTwo-sided p-value
OursCHOIS195 / 22586.7%8.01 × 10⁻³¹
Ours + SDFCHOIS + SDF182 / 21684.3%1.19 × 10⁻²⁵
OursCHOIS + SDF204 / 22192.3%6.92 × 10⁻⁴²
Ours + SDFCHOIS180 / 22281.1%1.53 × 10⁻²¹
OursInfBaGel195 / 22985.2%1.23 × 10⁻²⁸
Ours + SDFInfBaGel188 / 23081.7%3.05 × 10⁻²³

Q6: Inference speed and memory usage

Following InfBaGel, we report FPS as generated motion frames per second of wall-clock computation. Results on a single NVIDIA RTX 3090 are included in Appendix E.4, Table 7. All methods use approximately 6 to 9 GB of peak GPU memory. Runtime is mainly affected by guidance and optimization where required. InfBaGel achieves higher throughput through consistency-based few-step denoising.

MethodFPS ↑
CHOIS3.234
MAMHOI (Ours)2.933
InfBaGel24.039
HOSIG0.23
Reviewer_UjDhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the challenge of simultaneously ensuring interaction feasibility within a given environment and synthesizing realistic human-object motion during scene-aware human-object interaction generation. To this end, the authors propose the MAMHOI, which utilizes "motion affordance" as an intermediate interface to decouple scene feasibility reasoning from human-object motion synthesis, with each component learned from distinct data sources, human-scene interaction data and scene-agnostic human-object data, respectively. Evaluations conducted on the OMOMO validation set and across approximately 820 trials based on LINGO scenes demonstrate that the proposed method reduces scene penetration compared to baselines while maintaining high-quality motion.

Critical strengths
  • The authors propose a new decomposition paradigm that balances the feasibility of reasoning about interactions within an environment with the synthesis of realistic human-object motion, thereby eliminating the reliance on paired human-object-scene data.

  • This submission demonstrates that scene feasibility reasoning and human-object motion generation can be learned independently without the need for paired data and then effectively combined at inference time via a simple 2D interface. In Table 2, the model without any test-time guidance outperforms CHOIS (with SDF guidance) across all human–scene and object–scene metrics. This result is compelling.

  • The evaluation is comprehensive and yields strong results. Experiments on both benchmarks and user studies demonstrate that the proposed method outperforms the baselines.

Critical weaknesses
  • One concern is that the affordance bounding box might inadvertently reveal the trajectory to be generated. During training, this box is constructed from the ground-truth trajectory itself. Its position and extent directly encode the direction and magnitude of the motion. Consequently, the model might learn to simply "read" the trajectory from the bounding box, rather than generating it based on scene constraints, raising doubts about whether the model has truly acquired the ability to perform scene reasoning.

  • This submission uses affordance to reason about the feasibility of interactions within an environment. However, the authors do not discuss how an object's volume or shape influences its affordance, for instance, whether the affordances of a large object and a small object are the same. This is a concerning point.

  • Although the paper reports strong quantitative results, it lacks an in-depth analysis of the loss in flexibility resulting from 2D-constrained affordances, such as the inability to perform interactions that are feasible in 3D space, like stepping over low obstacles

  • In the evaluation, the authors use A* path planning as the input for affordance generation; however, they do not analyze the performance gains provided by these affordances compared to using the A* path (or its inflated corridor) directly as a constraint. Consequently, the ablation study intended to justify the components of their framework lacks clarity.

Questions for authors
  • Have the authors discussed or tested the model's performance when affordance is inaccurate, in order to analyze whether the model generates motion under scene constraints or merely overfits to the box shape and reads the trajectory off it?

  • Could the authors discuss whether, within the same scene and along the same path, Factor I generates the same affordances for objects of different sizes (e.g., a cup versus a dining table)? Does the size of an object have an effect on Factor I?

  • Could the authors discuss the contribution of L_rank? It only reduced the percentage of in-scene penetration frames from 17.34% to 14.99% (Table 4), but doubled the percentage of non-scene FIDs (Table 1). The authors did not discuss this trade-off.

2 comments

We appreciate your recognition of our factorization approach and your constructive feedback. We address your comments and questions below.

W1 & Q1: Trajectory leakage and robustness to inaccurate affordance

The affordance bounding box represents spatial support without explicitly encoding temporal order or a complete trajectory.At scene-aware inference, affordances are predicted from scene depth, the planned path, and text, without access to the target motion.

To test whether the generator relies on the box as a trajectory shortcut, we remove the affordance or replace it with an inflated A* corridor while keeping the generator and planned paths fixed. Both interventions preserve similar path errors but substantially increase human-scene penetration, suggesting that affordance contributes to scene-compatible execution beyond path control. The new results are reported in Appendix E.2, Table 4. Object-scene penetration depth is averaged per frame.

Spatial conditionHuman-scene penetration rate (%) ↓Object-scene penetration depth (×10⁻³) ↓Object-path error ↓
A* corridor25.9661.7940.1490
No scene condition30.2342.2130.1527
Original scene occupancy18.7251.7860.1500
Ours14.9901.4400.1492

W2 & Q2: Object geometry and spatial requirements

We acknowledge that Factor I does not account for object geometry, which may lead to information loss. Our factorization transfers coarse spatial support from HSI to HOI without requiring identical affordances across domains. The affordance is a soft condition, a shared support region can guide different interactions without requiring each motion to follow its boundaries strictly. Factor II then combines this soft condition with object geometry, the planned path, and the learned HOI prior to generate coordinated motion.

Our new per-object evaluation (Appendix E.2, Table 5) compares generation with and without predicted affordance guidance. Across all six objects, this guidance reduces human-scene penetration rates and object-scene penetration depths while maintaining similar path errors.

Entries show Without → With (relative change). Affordance is removed only at inference, with the generator and paths fixed. These results provide empirical evidence that HSI-learned affordances offer useful guidance for HOI generation across different objects, although they do not establish object-specific clearance.

ObjectHuman-scene penetration rate (%) ↓Object-scene penetration depth (×10⁻³) ↓Object-path error ↓
clothesstand31.084 → 12.200 (-60.75%)3.590 → 1.193 (-66.77%)0.1527 → 0.1482 (-2.95%)
floorlamp29.291 → 13.279 (-54.67%)1.651 → 1.535 (-7.03%)0.1553 → 0.1511 (-2.70%)
largebox28.641 → 14.312 (-50.03%)1.666 → 0.944 (-43.34%)0.1530 → 0.1482 (-3.14%)
largetable32.180 → 16.164 (-49.77%)3.425 → 3.152 (-7.97%)0.1487 → 0.1471 (-1.08%)
smallbox26.518 → 15.657 (-40.96%)1.432 → 0.511 (-64.32%)0.1562 → 0.1544 (-1.15%)
whitechair33.691 → 18.327 (-45.60%)1.512 → 1.303 (-13.82%)0.1505 → 0.1463 (-2.79%)

We have clarified this limitation in lines 508 to 516 and will explore object-aware affordances in future work.

W3: Height-dependent feasibility

We agree that 2D affordances lose height-dependent information, limiting interactions such as stepping over low obstacles. This is a limitation of the current representation, not a fundamental limitation of our factorization. We have clarified this in lines 508 to 516 and will explore richer height-aware representations in future work.

W4: Learned affordance versus A* corridors

Without affordance conditioning during training, the generator reduces to the CHOIS baseline, which is already included in our comparisons. To further isolate the contribution of the learned affordance within our trained model, we also test removing it or replacing it with an A* corridor or raw scene occupancy at inference, keeping the generator and planned paths fixed. These variants yield higher scene penetration despite similar path errors (W1 & Q1, Table 4), showing that learned affordance provides useful spatial guidance beyond the planned path alone.

Q3: Ranking loss and motion fidelity

L_rank encourages a reweighting of the conditional motion distribution toward motions that better match the affordance. This helps prevent the generator from ignoring the affordance condition, but also shifts its output distribution away from the broader distribution learned from scene-absent HOI data. Since FID in Table 1 is evaluated against scene-absent OMOMO motion, this shift can increase FID even when scene compatibility improves.

Even without L_rank, most penetration in relatively simple cases has already been avoided, leaving failures dominated by particularly challenging scene configurations. Adding L_rank reduces the penetration-frame rate from 17.34% to 14.99%, corresponding to a 13.6% relative reduction in the remaining penetration. We have added a discussion of this trade-off and the role of L_rank in Appendix E.3 of the revised paper (lines 1127 to 1133).

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

This paper proposes MAMHOI, a method for generating scene-aware 3D HOI without relying on paired HOSI data. MAMHOI decomposes the synthesis process through an intermediate affordance map interface: the scene-conditioned image network takes scene depth, planned object trajectories, and text as input, outputs a 2D task-specific affordance map, and then lifts it to a local occupancy voxel representation. Based on these localized affordance features, the diffusion backbone generates synchronized human and object motions. These two stages are trained independently on the LINGO (human-scene) and OMOMO (human-object) datasets. The results show that the scene penetration is reduced while the interaction quality is preserved.

Critical strengths
  1. This method addresses the scarcity of paired human-object-scene data. The scene-understanding model and the HOI motion generator are trained separately on LINGO and OMOMO, without requiring paired human-object-scene interaction data. The idea of using complementary data sources to address practical data constraints is insightful.

  2. The counterfactual ranking objective (LrankL_{\mathrm{rank}}) is a interesting component. Pairing motions with similar endpoint configurations but different trajectories can mitigate the gradient vanishing problem of standard SDF-based anchors within the feasible region, thereby providing a continuous relative signal to guide the selection of navigation modes along alternative obstacle‑avoidance paths.

Critical weaknesses
  1. The affordance representation may be insufficient for complex 3D feasibility. The method predicts a 2D affordance map, and then lifts it to 3D. This may lose height-dependent feasibility differences, such as those around overhanging structures, semi-open cabinets, or under-table spaces. In addition, affordance labels are constructed from relaxed boxes around motion trajectories, so it is unclear whether the learned regions correspond to truly feasible space. The paper does not evaluate these types of 3D structures, limiting the evidence for the method’s scene-feasibility claims.

  2. The central assumption that motion affordances transfer across datasets to human-object interaction is not sufficiently validated. The Scene Understanding Model is trained on human-scene data from LINGO, while the motion generator is trained on human-object data from OMOMO. The scene model takes scene depth, object path, and text as input, but not object geometry, so it is unclear whether it predicts interaction space appropriate for objects of different sizes and shapes. Although this separate training avoids the need for paired human-object-scene data, the paper does not directly validate whether motion affordances learned from LINGO transfer reliably to object-centric motion trajectory. This weakens the evidence for the proposed factorization.

Questions for authors
  1. Validity of Cross-Dataset Transfer and Lack of Object Geometry Awareness. How is the "planner-provided object path" instantiated during training on LINGO? Given the absence of movable large objects in LINGO, how is this path constructed and aligned with the affordance labels derived purely from human body trajectories?

  2. Potential Optimization Conflict Between LrankL_{\mathrm{rank}} and Motion Fidelity. The primary learning objective of the motion diffusion model is reconstruction and kinematic consistency, which fits the ground-true human motion samples, whereas LrankL_{\mathrm{rank}} imposes a heuristic discriminative margin on synthetic counterfactual pairs. In Table 1 on the scene-free OMOMO validation split, removing LrankL_{\mathrm{rank}} dramatically improves FID from 1.8361 to 0.9403. This suggests an optimization trade-off or gradient conflict between LrankL_{\mathrm{rank}} and the native motion reconstruction objectivel. Could the authors clarify the underlying trade-off between these two objectives? This directly impacts the assessment of LrankL_{\mathrm{rank}} as a key technical contribution.

3 comments

Thank you for recognizing the value of learning from complementary data sources. We address your comments and questions below.

W1: 2D affordance limitations and supervision

We agree that 2D affordances lose height-dependent information, limiting complex 3D scene reasoning. This is a limitation of the current representation, not a fundamental limitation of our factorization. We have clarified this in the revised paper (lines 508 to 516) and will explore richer height-aware representations in future work.

We agree that using relaxed bounding boxes as supervision biases the learned affordance toward the motion distribution in the training data and it can omit feasible regions that are rarely used. However, since the motion generator is also trained with affordance conditions derived from observed motion, we expect the learned affordance to better match the spatial guidance it receives during training. To test its effect on generation, we added an ablation (Appendix E.2, Table 4) that replaces the learned affordance with the original scene occupancy while keeping the generator and planned object paths fixed. The results below show reduced human-scene penetration rates and object-scene penetration depths, averaged per frame.

Spatial conditionHuman-scene penetration rate (%) ↓Object-scene penetration depth (×10⁻³) ↓Object-path error ↓
A* corridor25.9661.7940.1490
No scene condition30.2342.2130.1527
Original scene occupancy18.7251.7860.1500
Ours14.9901.4400.1492

W2: Affordance transfer across datasets and objects

We acknowledge that the current predictor does not account for object geometry, which may lead to information loss. In our factorization design, we transfer coarse spatial support from HSI to HOI without requiring identical affordances across domains. The affordance acts as a soft condition, which the generator combines with object geometry, the planned path, and the learned HOI prior to produce coordinated motion.

Our new per-object evaluation (Appendix E.2, Table 5) measures the gain in scene awareness for each object by comparing generation with and without predicted affordance guidance. Across all six objects, this guidance reduces human-scene penetration rates and object-scene penetration depths while maintaining similar path errors. These results support the usefulness of the transferred guidance, although object-specific clearance is not guaranteed.

Entries show Without → With (relative change). Affordance is removed only at inference, with the generator and paths fixed. Object-scene depth is averaged per frame.

ObjectHuman-scene penetration rate (%) ↓Object-scene penetration depth (×10⁻³) ↓Object-path error ↓
clothesstand31.084 → 12.200 (-60.75%)3.590 → 1.193 (-66.77%)0.1527 → 0.1482 (-2.95%)
floorlamp29.291 → 13.279 (-54.67%)1.651 → 1.535 (-7.03%)0.1553 → 0.1511 (-2.70%)
largebox28.641 → 14.312 (-50.03%)1.666 → 0.944 (-43.34%)0.1530 → 0.1482 (-3.14%)
largetable32.180 → 16.164 (-49.77%)3.425 → 3.152 (-7.97%)0.1487 → 0.1471 (-1.08%)
smallbox26.518 → 15.657 (-40.96%)1.432 → 0.511 (-64.32%)0.1562 → 0.1544 (-1.15%)
whitechair33.691 → 18.327 (-45.60%)1.512 → 1.303 (-13.82%)0.1505 → 0.1463 (-2.79%)

We have clarified this limitation in lines 508 to 516 and will explore object-aware affordances in future work.

Q1: Path construction for LINGO training

We have added the path-construction details in Appendix D.1 of the revised paper (lines 806 to 808). During Factor 1 training with Lingo data, we randomly sample pseudo object start and end positions near the corresponding pelvis endpoints and re-plan an A* path between them in the same scene. The resulting path remains aligned with the observed human transition and its affordance label. This augmentation is intended to reduce the conditioning gap between HSI training and path-conditioned HOI inference.

Q2: Ranking loss and motion fidelity

L_rank encourages a reweighting of the conditional motion distribution toward motions that better match the affordance. This helps prevent the generator from ignoring the affordance condition, but also shifts its output distribution away from the broader distribution learned from scene-absent HOI data. Since FID in Table 1 is evaluated against scene-absent OMOMO motion, this shift can increase FID even when scene compatibility improves.

Even without L_rank, most penetration in relatively simple cases has already been avoided, leaving failures dominated by particularly challenging scene configurations. Adding L_rank reduces the penetration-frame rate from 17.34% to 14.99%, corresponding to a 13.6% relative reduction in the remaining penetration. We have added a discussion of this trade-off and the role of L_rank in Appendix E.3 of the revised paper (lines 1127 to 1133).

Reviewer_uHGKOpenReview ↗

Official Comment by Reviewer uHGK

Thank you for your reply, your additional explanation and experimental response have addressed most of my concerns. Although inherent limitations still exist, the comparative experiment has demonstrated the usefulness of the method for these limitations. I suggest the author add a limitation section in the paper, and I will maintain my score.

Many thanks for your response. We sincerely appreciate your constructive feedback and your time in reviewing our work. We have included a paragraph discussing the limitations in the Discussion section of the revised manuscript.

Comments on the paper

1 comment
Area_Chair_83kBOpenReview ↗

Invitation to Participate in the Discussion

The authors have provided their responses to the reviewers' comments. We would appreciate it if the reviewers could review the responses and share any further thoughts during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.