acceptodds
Under review as a conference paper at ICLR 2027

Harnessing Multimodal Large Language Models for Training-free Human-object Interaction Detection

Abstract

Human-object interaction (HOI) detection aims to localize human-object pairs and recognize their interactions. Traditional supervised methods perform strongly but rely on task-specific training. Recent multimodal large language models (MLLMs) offer a promising route to training-free HOI detection through their broad visual-semantic knowledge. However, existing approaches leave semantic reasoning and visual perception loosely coordinated, inducing severe perceptual passivity where early semantic assumptions bias relation grounding and lead to self-reinforcing semantic circularity. To address these limitations, we propose HarnessHOI, a training-free framework that transforms passive MLLM interpretation into active interaction-guided perceptual reasoning harness. Central to HarnessHOI is an interaction-guided perception mechanism that proactively projects emerging interaction hypotheses back into visual space, actively discovering instances and refining ambiguous evidence through targeted observation. Subsequent relation-agnostic geometric adjudication reconciles these multi-source hypotheses under persistent object identities, establishing a unified visual grounding that decouples final predictions from initial proposal biases. Extensive experiments on HICO-DET and V-COCO show that HarnessHOI establishes state-of-the-art results among training-free HOI methods, demonstrating the effectiveness of the proposed MLLM harnessing strategy for complex visual-structuring tasks. The code will be released upon the publication.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Human object interaction detection, Multimodal Large Language Model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_W6rYOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HarnessHOI, a training-free human-object interaction (HOI) detection framework that integrates perception and interaction reasoning using a single frozen MLLM without auxiliary grounding models.

The framework consists of three main stages. First, interaction-guided perception performs initial participant localization, followed by hypothesis-directed discovery to identify potentially missed participants. When additional visual evidence is needed, targeted visual observation examines selected image regions to recover or refine participant groundings. Second, relation-agnostic geometric adjudication selects a consistent bounding box for each participant independently of the proposed interactions. Third, grounded interaction reasoning assesses human-object pairs and verifies interaction relations independently of the initial interaction proposals.

The method is evaluated on HICO-DET and V-COCO, including open-vocabulary settings. The authors further provide component ablations, controlled comparisons of design choices, and analyses of grounding coverage and inference cost to assess the proposed framework.

Critical strengths

S1. Unified Interaction-Guided Perception and HOI Reasoning with a Single Frozen MLLM

The paper proposes a unified framework that integrates participant localization, interaction-guided discovery, targeted visual observation, and HOI reasoning using a single frozen MLLM, without task-specific fine-tuning or auxiliary grounding models. This structured perception–reasoning pipeline achieves competitive training-free performance on HICO-DET (33.20 Full mAP) and V-COCO (44.65/49.70 S1/S2 role AP), demonstrating the effectiveness of the overall framework.

S2. Explicit Separation of Participant Grounding and Interaction Verification

The framework separates participant grounding from interaction reasoning through relation-agnostic geometric adjudication and proposal-independent relation verification. This design reduces reliance on initial interaction hypotheses when selecting participant boxes and assessing relations. Table 4 provides empirical support for these design choices, showing improvements of 0.41 and 0.70 mAP, respectively, over their controlled alternatives on HICO-DET.

Critical weaknesses

W1. Insufficient Direct Validation of the Effectiveness of Interaction-Guided Perception and Participant Recovery

A central contribution of this paper is interaction-guided perception, which employs hypothesis-directed discovery and targeted visual observation to recover participants missed during initial localization and improve participant grounding. However, the empirical evidence is insufficient to establish how accurately and reliably these mechanisms recover failures of initial localization.

According to Table 25 in Appendix D.3, hypothesis-directed discovery increases grounding coverage by 571 interaction annotations on HICO-DET and 217 role annotations on V-COCO, whereas subsequent targeted visual observation contributes only 12 and 23 additional annotations, respectively. Furthermore, Table 31 in Appendix E reports an average of only 0.19 and 0.40 local observations per image on the two benchmarks.

The authors provide qualitative examples of recovering initially missed objects in Appendix F and demonstrate in Table 4 that interaction-guided discovery and targeted visual observation outperform repeated participant localization and non-directed observation, respectively. These results provide evidence supporting the effectiveness of the proposed design choices. However, the paper does not directly quantify how accurately the MLLM-proposed observation regions contain their intended targets or how frequently these observations successfully recover participants missed during initial localization.

In particular, the grounding coverage reported in Table 25 is measured at the interaction-annotation level. Consequently, it does not clearly distinguish improvements arising from the recovery of previously missed unique human/object instances from those resulting from bounding-box refinement of already localized participants.

Moreover, Table 3 shows that Direct HOI generation achieves 15.62 mAP on HICO-DET, whereas the Initial-instance baseline, which performs visual pair assessment and relation verification over initially localized participants, achieves 31.45 mAP. The complete HarnessHOI framework, which additionally incorporates interaction-guided perception and relation-agnostic geometric adjudication, achieves 33.20 mAP, representing a further improvement of 1.75 mAP over the Initial-instance baseline.

Although the difference between Direct HOI generation and the Initial-instance baseline cannot be attributed exclusively to relation verification, since their candidate generation and assessment procedures also differ, these results highlight the importance of more directly examining the contribution of the additional perceptual stages.

Overall, the experiments demonstrate that interaction-guided perception improves final HOI detection performance. Nevertheless, the extent to which hypothesis-directed discovery and targeted visual observation effectively resolve specific perceptual failures, particularly the recovery of initially missed participants, remains insufficiently characterized. This limits the evidence supporting the practical effectiveness and significance of a central methodological contribution.

W2. Insufficient Evidence of Methodological Distinction from AgentHOI and the Independent Benefits of the Single-MLLM Design

This paper proposes a framework that uses a single frozen MLLM to perform interaction-guided perception, relation-agnostic geometric adjudication, and grounded interaction reasoning.

However, AgentHOI (Lei et al., 2026) already employs multi-round HOI re-mining to discover interactions missed during initial inference and uses instance-specific descriptions together with GroundingDINO to localize interaction participants. Although the two methods differ in their implementations, they share the underlying idea of leveraging previously inferred interactions to guide the discovery of additional interactions or participants.

HarnessHOI differs from AgentHOI by performing participant localization and relation verification using the same MLLM without an auxiliary grounding model, while introducing relation-agnostic geometric adjudication and grounded interaction reasoning. Nevertheless, the independent methodological advantages and significance of these components relative to existing approaches that combine multi-round interaction reasoning with specialized grounding models remain insufficiently established.

The authors provide controlled comparisons of individual design choices in Table 4 and re-evaluate the released AgentHOI predictions using a common evaluator in Appendix C.3. These experiments offer useful evidence for the effectiveness of HarnessHOI. However, the comparisons in Table 4 evaluate alternative configurations within the proposed framework, while neither Direct HOI generation nor repeated participant localization directly reproduces the multi-round HOI re-mining procedure employed by AgentHOI. Furthermore, the common-evaluator comparison with AgentHOI involves different foundation models and localization architectures, making it difficult to attribute the observed performance differences specifically to the newly proposed components of HarnessHOI.

An alternative design would preserve the MLLM-generated interaction hypotheses and target descriptions while delegating participant localization to a frozen specialized grounding model, such as GroundingDINO. Such a configuration would retain the general objective of interaction-guided participant discovery. However, the paper does not directly examine whether using the same MLLM for participant localization and targeted visual observation provides advantages in localization accuracy or computational efficiency over this alternative.

In this regard, Table 31 in Appendix E reports inference costs of 13.26 GPU-worker seconds per image on HICO-DET and 22.07 GPU-worker seconds per image on V-COCO. These measurements represent GPU-worker computation rather than per-image wall-clock latency, but they are relevant to assessing the computational overhead of the proposed framework. Although the authors provide performance–cost comparisons against repeated Direct HOI generation, the absence of a controlled comparison with a specialized grounding-based alternative leaves the accuracy–computation trade-off of the single-MLLM design insufficiently characterized.

The ability to perform the entire pipeline using a single frozen MLLM without auxiliary grounding models is a meaningful advantage. Nevertheless, given that AgentHOI already incorporates interaction-guided re-mining and instance grounding, the paper does not sufficiently establish what distinct methodological advances are provided by its localization, geometric adjudication, and relation verification procedures relative to the closest prior approach. This uncertainty is consequential to my assessment of the paper's novelty and significance.

Questions for authors

Q1. Effectiveness of Interaction-Guided Perception in Recovering Missed Participants

A central contribution of this paper is interaction-guided perception, which employs hypothesis-directed discovery and targeted visual observation to recover participants missed during initial localization and improve participant grounding.

The authors provide stage-wise grounding coverage analyses in Appendix D.3 (Tables 25–27) and evaluate targeted visual observation against non-directed observation in Table 4. However, Table 25 shows that targeted visual observation increases grounding coverage by only 12 interaction annotations on HICO-DET and 23 role annotations on V-COCO. Furthermore, Table 31 reports an average of only 0.19 and 0.40 local observations per image, respectively. These results raise questions about how effectively the proposed observation mechanism addresses failures in initial localization.

I would appreciate clarification on the following points:

(1) For executed targeted visual observations where the intended target can be identified from ground-truth annotations, what proportion of the proposed observation regions actually contain the intended target, and what is the subsequent localization success rate? In particular, how frequently does targeted visual observation successfully recover unique human or object instances that were missed during initial localization?

(2) Among the improvements in grounding coverage achieved through hypothesis-directed discovery and targeted visual observation, how much can be attributed to recovering previously missed participant instances, as opposed to refining the bounding boxes of already localized participants? Since the reported coverage is measured at the interaction-annotation level, it does not directly distinguish these two sources of improvement.

(3) Could the authors provide an oracle-grounding evaluation in which ground-truth human and object bounding boxes are supplied to the relation verification stage, to help distinguish limitations in participant localization from those in grounded interaction reasoning? Additionally, comparing the proposed observation regions against oracle observation regions derived from ground-truth target boxes would help isolate errors arising from region selection and subsequent localization. If these oracle evaluations are not feasible, an analysis based on the existing results that separates localization-related errors from relation reasoning errors would also be informative.

These questions aim to clarify which perceptual failures are actually addressed by interaction-guided perception and how effectively they are resolved. In particular, quantitative evidence on observation-region accuracy and the recovery of previously missed participant instances would materially strengthen the support for this central contribution.

Q2. Methodological Distinction from AgentHOI and the Value of Single-MLLM Localization

AgentHOI (Lei et al., 2026) already employs multi-round HOI re-mining to discover interactions missed during initial inference and uses GroundingDINO to localize interaction participants. HarnessHOI shares the general idea of using interaction hypotheses to guide further participant discovery, although the two methods differ in their implementations.

The authors provide controlled comparisons of individual design choices in Table 4 and re-evaluate the released AgentHOI predictions using a common evaluator in Appendix C.3. These experiments provide useful evidence for the effectiveness of HarnessHOI. However, the cross-method comparison involves different foundation models and localization architectures, making it difficult to isolate the contributions of the newly proposed components relative to the closest prior approach.

I would appreciate clarification on the following points:

(1) What specific methodological capabilities does the proposed hypothesis-directed discovery introduce beyond the multi-round HOI re-mining employed by AgentHOI? Furthermore, what distinct advantages do relation-agnostic geometric adjudication and grounded interaction reasoning provide over the grounding and interaction reasoning procedures used in AgentHOI? In particular, how do the controlled comparisons in Table 4 substantiate these distinctions relative to prior work, rather than only demonstrating improvements over alternative configurations within HarnessHOI?

(2) To assess the value of performing participant localization and targeted visual observation using the same frozen MLLM, would it be possible to conduct a controlled comparison in which the MLLM-generated interaction hypotheses and target descriptions are preserved, while participant localization is delegated to a frozen grounding model such as GroundingDINO? If feasible, comparing participant grounding accuracy, final HOI detection performance, and overall inference cost while keeping the downstream grounded interaction reasoning procedure unchanged would help clarify the benefits and accuracy–computation trade-offs of the single-MLLM design. If such a comparison is not feasible, could the authors identify the evidence in the current experiments that supports the advantages of MLLM-based localization over a specialized grounding model?

The purpose of these questions is not to require the incorporation of an external grounding model or a reduction in inference cost. Rather, it is to assess the distinct methodological contribution of HarnessHOI relative to AgentHOI. Clear empirical evidence supporting the proposed design choices and the value of single-MLLM localization would be important for evaluating the paper's novelty and significance.

11 comments
AuthorsOpenReview ↗

Response to Reviewer W6rY (1/6): Observation Regions and Localization

We sincerely thank you for recognizing the value of single-MLLM perception and reasoning, and for highlighting the explicit separation of participant grounding and interaction verification. We appreciate your detailed questions about observation-region accuracy, recovery of unique participants and the methodological distinction from AgentHOI. We have added request-level diagnostics, instance-recovery analyses, GT-input experiments and controlled localizer comparisons to provide direct evidence on these aspects of the framework.

W1. Insufficient Direct Validation of the Effectiveness of Interaction-Guided Perception and Participant Recovery

However, the paper does not directly quantify how accurately the MLLM-proposed observation regions contain their intended targets or how frequently these observations successfully recover participants missed during initial localization.

In particular, the grounding coverage reported in Table 25 is measured at the interaction-annotation level.

Q1(1). For executed targeted visual observations where the intended target can be identified from ground-truth annotations, what proportion of the proposed observation regions actually contain the intended target, and what is the subsequent localization success rate? In particular, how frequently does targeted visual observation successfully recover unique human or object instances that were missed during initial localization?

We measured both region accuracy and localization success on the two complete test sets. Appendix D.3 in the original paper reports coverage from discovery and observation, and we added the request-level diagnostics your question asks for in Table 6(a) of the revised PDF. Of the executed observations, 1,387/1,835 (HICO-DET) and 1,591/1,978 (V-COCO) can be matched uniquely to a ground-truth target, and we use them as the denominator. A localization counts as successful only when the instance is correct and the box reaches IoU ≥ 0.5. Equation (9) maps every box proposed inside a local crop back to full-image coordinates, so a region-level hit and an instance-level recovery are measured in the same frame and the rates below are directly comparable.

Request-level measureHICO-DETV-COCO
Region contains the target centre1,341/1,387 (96.7%)1,527/1,591 (96.0%)
Region contains ≥ half the target area1,228/1,387 (88.5%)1,417/1,591 (89.1%)
Region contains the entire target box862/1,387 (62.1%)1,119/1,591 (70.3%)
Qwen direct localization success1,012/1,387 (73.0%)1,104/1,591 (69.4%)
Direct success plus adopted corrections1,198/1,387 (86.4%)1,318/1,591 (82.8%)

The four direct outcomes are mutually exclusive and sum to the denominator.

Qwen direct localization outcomeHICO-DET (1,387)V-COCO (1,591)
Correct localization1,012 (73.0%)1,104 (69.4%)
Box deviation248 (17.9%)302 (19.0%)
Empty localization87 (6.3%)121 (7.6%)
Wrong same-class instance40 (2.9%)64 (4.0%)

Among requests where the instance was identified correctly but the box fell below IoU 0.5, observation corrections succeeded in 186/248 (75.0%) and 214/302 (70.9%) cases. Region proposals contain the target centre in 96.7% / 96.0% of uniquely matchable requests. The accepted corrections localize the residual error to the region-to-box step, which observation resolves in 75.0% / 70.9% of deviated cases. Wrong same-class selection occurs in 7.8% (40/512) / 9.4% (64/684) of requests with same-class distractors and in 0% (0/875, 0/907) without them. Deduplicating by image and physical instance gives the instance-level count: targeted observation recovers 0 (HICO-DET) and 5 (V-COCO) unique participants absent from the initial set under the submitted trigger, and 52 / 46 when the trigger is extended. Appendix D.3 gives the full per-stage decomposition. Appendix D.3 in the revised manuscript carries the request-level results and the four-way classification. These results quantify observation-region accuracy and participant recovery on both complete test sets, and show that targeted observation supplies usable visual evidence.

AuthorsOpenReview ↗

Response to Reviewer W6rY (2/6): Recovering Missed Instances versus Refining Boxes

Q1(2). Among the improvements in grounding coverage achieved through hypothesis-directed discovery and targeted visual observation, how much can be attributed to recovering previously missed participant instances, as opposed to refining the bounding boxes of already localized participants? Since the reported coverage is measured at the interaction-annotation level, it does not directly distinguish these two sources of improvement.

Recovery and refinement can now be separated at the instance level. Appendix D.3 of the submitted paper reports annotation-level coverage, and Equation (17) states what that quantity can and cannot establish: with M_c the number of annotations that have an eligible matching prediction and G_c the ground-truth count, one eligible prediction per covered annotation yields a class-wise AP bound of U_c(HICO) = (1/11) Σ_t 1[t ≤ M_c/G_c] and U_c(VCOCO) = M_c/G_c. Coverage bounds attainable AP under fixed boxes and labels, but it says nothing about whether a box came from a recovered participant or from a refined one. We have added the instance-level decomposition the question asks for, after deduplicating by image and physical instance and assigning each target to the stage that first brought new coverage.

StageConfirmed initially unlistedClass/box correction or initially ambiguous identityDeduplicated total
HICO-DET, hypothesis-directed discovery27442316
HICO-DET, targeted observation (submitted trigger)01212
V-COCO, hypothesis-directed discovery63063
V-COCO, targeted observation (submitted trigger)51419
Total, submitted policy34268410

The 12 newly covered HICO-DET annotations map to 12 unique targets, all involving class/box corrections or identity resolution. The 23 newly covered V-COCO annotations map to 19 unique targets: 5 initially unlisted participants and 14 class/box corrections or identity resolutions. Appendices D.3 and D.4 separately record the 479 / 424 instances whose final box came from observation. Under the submitted policy, discovery recovers 274/63 initially unlisted participants and observation recovers 0/5; observation also refines groundings and supplies local visual evidence.

With two additional triggers for already-boxed targets, at the same two-observation cap, observation recovers 52 / 46 such participants in total (52 / 41 additional over the submitted policy) and adds 140 / 179 covered annotations. Deduplicated observation targets rise to 75 / 103, of which 23 / 57 are class or box corrections. The two datasets together gain 147 new targets beyond the submitted policy (93 initially unlisted, 3 class corrections, 41 box corrections, 10 with initially ambiguous identity). Deduplicated targets from discovery and observation rise from 410 to 557. Table 6(b) in the revised PDF reports this variant as a trigger-sensitivity control, with the submitted policy retained for the main tables. Appendix D.3 provides the deduplication rule and the per-stage breakdown. We can now separate recovery of missed participants from refinement of boxes.

AuthorsOpenReview ↗

Response to Reviewer W6rY (3/6): Separating Localization and Relation Errors

Q1(3). Could the authors provide an oracle-grounding evaluation in which ground-truth human and object bounding boxes are supplied to the relation verification stage, to help distinguish limitations in participant localization from those in grounded interaction reasoning? Additionally, comparing the proposed observation regions against oracle observation regions derived from ground-truth target boxes would help isolate errors arising from region selection and subsequent localization. If these oracle evaluations are not feasible, an analysis based on the existing results that separates localization-related errors from relation reasoning errors would also be informative.

We added two GT-input conditions and a pair-reasoning ablation on both complete test sets. Table 4 in the submitted paper reports component controls; Appendix D.8 adds these diagnostics under the same matching and AP rules. Equation (6) scores each triplet as s(h,a,o) = q_theta(true | x_rel), where x_rel contains the selected boxes and visual evidence. O1 supplies GT observation regions while fixing every final output box and discarding observation-induced box updates. O2 initializes GT participants, supplies GT observation regions and re-runs downstream prediction; its final boxes are generated by that prediction process. The pair-reasoning ablation retains detection and single-target action classification while removing inference about specific human-target combinations.

ConditionHICO Default FullHICO KO FullV-COCO S1V-COCO S2
HarnessHOI33.2036.8544.6549.70
O1: GT observation regions, original output boxes held fixed36.6840.4248.0953.24
O2: GT participant initialization + GT observation regions, downstream re-run40.7245.3158.4662.18
Detection + single-target action classification, pair reasoning removed26.1829.3435.7240.18

In O1 we replace predicted observation regions with GT-derived ones, hold every final output box fixed, and discard observation-induced box updates. HICO Default Full rises by 3.48, KO Full by 3.57, and V-COCO S1/S2 by 3.44 / 3.54. With the boxes fixed, this gain can only come from visual evidence reaching relation verification. O2 changes the participant initialization and the resulting downstream candidates, raising HICO Default Full and KO Full by another 4.04 / 4.89 points over O1. Removing pair-specific reasoning reduces HICO Default/KO Full by 7.02 / 7.51 and V-COCO S1/S2 by 8.93 / 9.52, demonstrating the contribution of judgments about the specific human-target pair.

Scoring O1 under the original HICO evaluation implementation gives the same +3.48 / +3.57, with a constant 0.23 to 0.24 point offset between evaluators (full method 33.20 vs 32.97, O1 36.68 vs 36.45), and the O2−O1 gap becomes 4.12 / 4.95. O1's gains are identical across evaluation implementations; both preserve the additional improvement from GT participant initialization.

Appendix D.8 of the revised PDF reports the input replacement, box handling and full metrics for all three conditions. Together, these diagnostics quantify the effects of visual evidence, participant initialization and pair-specific reasoning on final detection.

Reviewer_nNkXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes HarnessHOI, an interaction-centric inference framework that utilizes frozen MLLM. By combining interaction-guided perception, relation-agnostic geometric adjudication, and grounded interaction reasoning, we achieve superior performance in HICO-DET and V-COCO compared to existing training-free methods

Critical strengths
  1. The structure connecting interaction-guided perception, geometric adjudication, and relation verification is intuitive and rational, with clearly defined roles for each component.
  2. Achieving an HICO-DET 33.20 mAP and a V-COCO 44.65/49.70 AP with a single frozen MLLM alone, without additional task-specific fine-tuning or an external object detector, demonstrates better performance than existing training-free methods.
  3. Well analyze the effectiveness of individual designs by comparing them not only with ablations that remove each component but also with various alternatives such as repeated localization, random observation, and relation-conditioned grounding.
Critical weaknesses
  1. Limited Novelty : Techniques such as interaction discovery, image cropping, bounding-box selection, and VQA-based verification have been used in previous studies; in particular, compared to AgentHOI, it appears to be closer to an inference pipeline that combines existing capabilities rather than offering new algorithmic contributions.
  2. Practical benefits of the training-free approach: In situations where performance is lower than that of supervised HOI detectors and inference costs are high, the practical value of training-free methods has not been sufficiently demonstrated. For example, as with ImagineAgent[1]'s CoT-based learning, it is necessary to validate downstream utilities by using the proposed framework as a teacher to generate reasoning traces or pseudo-labels and improve the performance of the student model.
  3. Little Component Gain : Most performance improvements over direct generation occur in initial-instance verification, while the additional enhancements to the newly proposed perception and geometry components result in only a +1.75 mAP increase over the HICO-DET baseline. Although stochastic generation with a temperature of 0.7 is used, there is no component-specific multi-seed ablation, so it is unclear whether the small performance improvement is reproduced reliably.
  4. Limited contribution of targeted observation : The additional ground coverage obtained through targeted observation amounted to only 12 annotations in HICO-DET and 23 in V-COCO, suggesting that its practical impact as a key contribution is limited.
  5. Concerns about generalization when using training data : Although using Frozen MLLM, HarnessHOI utilizes evaluation feedback from the entire training set as well as some training images when developing the prompt and inference policies. In particular, it is unclear whether annotations for the unseen class in open-vocabulary evaluations were excluded during the development process, so additional explanation regarding strict unseen generalization and fairness is needed.
  6. Contribution on harder task : It presents compositional semantic-role reasoning as the main contribution, but evaluates only on HICO-DET and V-COCO. I think generalized performance should be verified on benchmarks that include a wider range of semantic roles, such as SWiG-HOI.

[1] Yuan et al, What if Agents Could Imagine? Reinforcing Open-Vocabulary HOI Comprehension through Generation, arxiv2026

Questions for authors

see weakness

7 comments
AuthorsOpenReview ↗

Response to Reviewer nNkX (1/6): The Contribution Claim

We sincerely thank you for recognizing the coherent three-stage structure, the strong training-free results from a single frozen MLLM, and the controlled comparisons of individual design choices. We appreciate your questions on methodological novelty, reliability, downstream utility and generalization. We have added same-backbone controls, paired three-seed ablations, a teacher-student experiment and an evaluation on SWiG-HOI, and clarified the open-vocabulary development protocol. We first explain the contribution and its distinction from AgentHOI.

W1. Limited Novelty : Techniques such as interaction discovery, image cropping, bounding-box selection, and VQA-based verification have been used in previous studies; in particular, compared to AgentHOI, it appears to be closer to an inference pipeline that combines existing capabilities rather than offering new algorithmic contributions.

HarnessHOI enables one frozen 9B MLLM to perform participant perception and grounded interaction reasoning without an auxiliary grounding model. Interaction-guided perception turns emerging interaction hypotheses into participant localizations and targeted observations. Relation-agnostic geometric adjudication then selects a shared grounding for each instance, and grounded interaction reasoning assesses multiple actions and semantic roles using these groundings and acquired views. This organization transforms passive MLLM inference into an active interaction-centric harness: hypotheses guide visual exploration, while final box selection and relation verification use instance-level visual evidence without the earlier relation proposal. Table 4 evaluates discovery and observation under matched budgets, and geometry and verification with their respective candidates and views fixed. Equations (1) and (2) state what is shared across them: D_theta(I, E_0, T) returns interaction hypotheses, newly localized participants and observation requests in one generation, and q_theta(l | x) = exp(z_l)/sum_j exp(z_j) normalizes every downstream answer over the same frozen model's candidate logits. There is no separate detector state and no separate scoring model. These controls test how interaction guidance and proposal-independent decisions affect detection. AgentHOI expands interaction triplets by multi-round re-mining and then gets participant boxes from GroundingDINO using instance descriptions, so its reasoning loop and its localizer keep separate states. Its paper lists GPT-4o or Qwen2.5-VL-72B with that grounding model. We instead expand the participant set: discovery recovers 274 HICO-DET and 63 V-COCO unique instances that were not listed initially, and the resulting boxes stay inside the same model's reasoning chain, shared by every later action and role judgment.

Once localization is inside the loop, early hypotheses become a liability as well as a resource. For this reason we select boxes without showing the relation proposal, gaining 0.41 / 0.47 / 0.54 HICO, V-COCO S1 and S2 points over a relation-conditioned control with the same candidate boxes and image views. We verify relations without the early proposal, gaining 0.70 / 0.73 / 0.86 with candidates, boxes and views fixed. Matched-budget alternatives give the same reading: hypothesis-directed discovery gains 0.75 / 0.78 / 0.82 over repeated localization, and targeted observation gains 0.50 HICO points over non-directed observation at equal observation count and region size. Each control changes one design choice under the matched conditions specified above. When localization is an external service, this comparison cannot be run.

With the same frozen Qwen3.5-9B, direct generation gives 15.62 mAP and the harness gives 33.20. On an identical backbone we obtain 4.92 points more than a locally adapted AgentHOI, and under a common evaluator the released GPT-4o + GroundingDINO predictions give 29.68 against our 33.20. Tables 1 and 2 of the original manuscript already report 33.20 HICO Default Full and 44.65 / 49.70 V-COCO S1/S2 as the strongest training-free results, with Unseen AP of 40.15 / 33.64 / 35.45 / 33.87 across the four open-vocabulary partitions. Appendix D.6 reports 32.11 / 43.60 / 48.40 on InternVL3-8B and 31.61 / 43.10 / 47.90 on Qwen2.5-VL-7B under the same configuration. Tables 3 and 4 report the component controls from 31.45 to 33.20 HICO Default Full and from 42.90 / 47.20 to 44.65 / 49.70 V-COCO S1/S2. Appendices C.3, D.7 and D.8 of the revision add the whole-pipeline controls. Taken together, the controls show that one frozen 9B MLLM runs the full pipeline without a separate grounding model: discovery and observation acquire participants and local evidence, geometric adjudication supplies shared instance groundings, and grounded verification scores their actions and roles using the same frozen model.

AuthorsOpenReview ↗

Response to Reviewer nNkX (2/6): Downstream Utility as a Teacher

W2. Practical benefits of the training-free approach: In situations where performance is lower than that of supervised HOI detectors and inference costs are high, the practical value of training-free methods has not been sufficiently demonstrated. For example, as with ImagineAgent[1]'s CoT-based learning, it is necessary to validate downstream utilities by using the proposed framework as a teacher to generate reasoning traces or pseudo-labels and improve the performance of the student model.

We followed the teacher–student route in W2 and measured label quality with a count-matched control. Table 3 of the original manuscript shows that the components change which candidates exist: with the initial instances fixed, the baseline reaches 31.45 HICO Default Full and 42.90 / 47.20 V-COCO S1/S2, while the full method reaches 33.20 and 44.65 / 49.70. Appendix H in the revised PDF tests whether those candidates also carry useful supervision. Following the pseudo-label route in your suggestion, we set up two teachers, the initial-instance baseline and the full HarnessHOI, over the same 2,048 HICO-DET training images. UPT-R50 with a frozen COCO-pretrained detector serves as the student. All conditions share interaction-head initialization, training images and 20,000 effective updates across seeds 66/67/68. We supervise only relations that were asked about and judged unambiguously, and we report both the raw label count and a per-image count-matched condition that equalizes positive and non-positive labels.

Pseudo-label conditionTeacherStudent Default FullStudent KO Full
Raw countInitial-instance baseline24.04 ± 0.3027.17 ± 0.26
Raw countHarnessHOI26.62 ± 0.2730.14 ± 0.32
Count-matchedInitial-instance baseline24.02 ± 0.2427.15 ± 0.22
Count-matchedHarnessHOI27.15 ± 0.3730.77 ± 0.18
Paired gain, HarnessHOI teacher minus baselineDefault FullDefault RareKO FullKO Non-Rare
Raw count+2.58 ± 0.22+0.42 ± 0.28+2.97 ± 0.19+3.45 ± 0.33
Count-matched+3.13 ± 0.17+0.58 ± 0.31+3.62 ± 0.22+4.17 ± 0.32

Matching the label count isolates label quality produced by discovery, observation and geometric adjudication. The gain is largest on Non-Rare and smaller on Rare. Teacher inference runs offline and student prediction runs UPT-R50, so the per-image cost of generating labels is paid once rather than at every student prediction. The student is a supervised detector, but its labels come from the frozen model rather than from ground-truth interaction annotations. With the label count matched, HarnessHOI adds +3.13 ± 0.17 Default Full and +4.17 ± 0.32 KO Non-Rare over the initial-instance teacher. This shows that discovery, observation and geometric adjudication supply a student with usable training signal from a single frozen model, and that the training-free framework has a concrete downstream use beyond its own detection AP.

AuthorsOpenReview ↗

Response to Reviewer nNkX (3/6): Component Gains across Three Seeds

W3. Little Component Gain : Most performance improvements over direct generation occur in initial-instance verification, while the additional enhancements to the newly proposed perception and geometry components result in only a +1.75 mAP increase over the HICO-DET baseline. Although stochastic generation with a temperature of 0.7 is used, there is no component-specific multi-seed ablation, so it is unclear whether the small performance improvement is reproduced reliably.

Yes, the gains reproduce across seeds. Table 3 in the submitted paper reports the component ablations under one seed. We have added paired three-seed runs on both complete test sets, sharing each seed's initial localization between a control and the full method (temperature 0.7, seeds 66/67/68), and report them in the new Table 5. Each gain comes from a paired per-seed difference, with ± as the sample SD of those differences.

Full method minus controlHICO Default FullHICO KO FullV-COCO S1V-COCO S2
Initial-instance baseline+1.72 ± 0.21+1.84 ± 0.17+1.71 ± 0.11+2.46 ± 0.10
Excluding discovery-added participants+0.68 ± 0.16+0.67 ± 0.14+0.66 ± 0.10+0.79 ± 0.08
Local observation disabled+0.40 ± 0.08+0.46 ± 0.10+0.32 ± 0.10+0.49 ± 0.13
First candidate box only+0.28 ± 0.07+0.32 ± 0.08+0.25 ± 0.06+0.24 ± 0.05
Relation proposal at box selection+0.44 ± 0.08+0.49 ± 0.07+0.45 ± 0.05+0.52 ± 0.04
Relation proposal at verification+0.74 ± 0.18+0.73 ± 0.19+0.76 ± 0.09+0.83 ± 0.05

All six controls have positive mean gains on all four metrics, and every gain exceeds its paired SD. The smallest is +0.24 ± 0.05 on V-COCO S2. Each control is paired with the full method on the same seed, so the reported mean and SD are within-seed statistics. Direct generation itself is stable at 15.66 ± 0.13 on HICO and 16.63 ± 0.24 / 21.62 ± 0.38 on V-COCO across the same seeds (15.62 in the main table is the seed-66 value), so the gap reflects the pipeline rather than sampling variation. The main table reports 33.20 for seed 66, while the three-seed means are 32.95 ± 0.22 on HICO Default Full and 44.65 ± 0.15 / 49.67 ± 0.10 on V-COCO S1/S2. The headline +1.75 reappears as the +1.72 ± 0.21 paired gain. Gains come from unrounded per-seed scores, so displayed means can differ by 0.01 from the paired values. Table 5 of the revised PDF reports the paired three-seed runs, and Appendix E.5 tabulates the absolute means. They show that every component contributes a reproducible gain on every metric.

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

This paper proposes HarnessHOI, a training-free framework for human-object interaction (HOI) detection using a frozen multimodal large language model (MLLM). It integrates three components: interaction-guided perception for discovering missing participants and requesting targeted visual observations, relation-agnostic geometric adjudication for selecting consistent instance-level localizations, and grounded interaction reasoning for verifying actions and semantic roles. Experiments on HICO-DET and V-COCO show improvements over existing training-free methods, with additional evaluations under open-vocabulary settings and component ablations.

Critical strengths

The paper identifies the coupling between interaction reasoning and visual perception as a central challenge. Its proposed pipeline explicitly connects interaction hypotheses, participant discovery, geometric grounding, and relation verification. The authors evaluate the method on two established benchmarks, HICO-DET and V-COCO, and report both standard and open-vocabulary results.

Critical weaknesses
  1. The object localization process needs a clearer explanation. Since the framework relies on the multimodal model to generate object locations, the quality of the initial bounding boxes may have a substantial impact on the final results.
  2. The improvements from individual components are relatively modest. The paper reports an increase in HICO-DET mAP from 31.45 to 33.20 and improvements on V-COCO from 42.90 to 44.65 and from 47.20 to 49.70. More detailed analysis would help establish which components contribute most to the overall performance.
  3. The computational cost is not sufficiently clear. The framework involves multiple perception and reasoning steps. Reporting inference time, model calls, and computational requirements would help readers assess its practical value compared with existing approaches.
Questions for authors
  1. How are the initial object detections and bounding boxes obtained? Please clarify the prompts, coordinate representation, post-processing procedures, and handling of missing or incorrectly localized objects.
  2. Do the proposed method and the compared methods use the same source of object detections? If they use different detectors or localization procedures, how do the authors ensure a fair comparison?
  3. Can the authors provide controlled experiments that isolate the contribution of the proposed perception and reasoning framework from the effects of the initial object localization process?
  4. What is the computational cost of the proposed framework in terms of inference time and model calls? How does this compare with existing agent-based HOI detection methods?
  5. What are the main failure cases of HarnessHOI? In particular, how does the framework perform when the initial bounding boxes are inaccurate or when multiple objects could plausibly correspond to the same interaction?
10 comments
AuthorsOpenReview ↗

Response to Reviewer woiC (1/5): Initial Localization and Post-Processing

We sincerely thank you for the careful review and for recognizing our framework's connection between interaction reasoning and visual perception, as well as the paper's clarity and soundness. We appreciate your questions about initial localization, comparison fairness, component contributions and computational cost. We have clarified the localization procedure and added same-backbone and localizer controls, paired component ablations and detailed cost analyses.

W1. The object localization process needs a clearer explanation. Since the framework relies on the multimodal model to generate object locations, the quality of the initial bounding boxes may have a substantial impact on the final results.

Q1. How are the initial object detections and bounding boxes obtained? Please clarify the prompts, coordinate representation, post-processing procedures, and handling of missing or incorrectly localized objects.

Section 3.2 in the submitted paper describes interaction-guided perception. Its starting boxes come from one full-image generation call to the frozen Qwen3.5-9B. No detector is used in the method itself. Initial boxes use full-image xyxy coordinates normalized to [0, 1000]. Each instance receives an identifier, so boxes proposed at different stages can be associated with the same person or object. Exact class-and-coordinate matches merge into one instance. Fully duplicated detections keep the highest score. We retain continuous scores and deduplicate the output exactly, with no NMS and no score threshold. Equation (9) maps a point (x, y) normalized to 0–1000 inside a crop [L, T, R, B] to the image coordinate (L + (R-L)x/1000, T + (B-T)y/1000), so boxes proposed in local views and in the full image share one coordinate frame before deduplication. Equation (13) scores direct generations by the geometric mean of their label-token probabilities, while every verification stage uses the normalized answer support q_theta(l | x) = exp(z_l)/sum_j exp(z_j) of Equation (2). The same normalized support is used for every downstream geometry, pair, action and binding question, so one frozen model produces every score in the pipeline.

Appendix G.1 of the revised PDF gives the localization prompt in full, together with the output structure. Conditioned on the full image and the public object vocabulary, it asks the model to localize "the people involved in visible interactions and all identifiable targets associated with them", to list each person followed by their targets, and to prioritize interaction participants while retaining nearby identifiable non-interacting targets when relevant. It requires one tight box per instance covering its full visible extent, with separate boxes for people and objects, and forbids inferring hidden instances or inventing boxes from coordinate patterns. The output is a compact JSON array with fields in the order bbox_2d, label, reference: boxes are full-image xyxy coordinates normalized to [0, 1000], labels come from the supplied vocabulary, and the reference describes visible appearance and position. The prompt excludes actions, instance IDs, confidence scores, explanations and Markdown, and the output is an empty array when no instance is identifiable. Instance identities are assigned after localization, when exact category-and-box duplicates are merged. Appendices G.2 to G.5 give the discovery, observation, geometry and relation-question prompts.

The same single-model inference loop handles missing and badly localized objects. Discovery adds participants implied by a partially formed interaction hypothesis. When a target has not been found or its boundary is unclear, the model can request a local view, at most two per image, and the resulting boxes enter that instance's candidate pool. Geometric adjudication then looks at the image and picks one box per instance based on its identity, class and visible extent, without seeing the earlier relation proposal. Downstream stages reuse that box for every action and role involving the instance. Appendix A.2 states the input, output and coordinate convention in one place, and Appendix B.3 gives the post-processing rules. The same frozen model handles both the initial localization and every later perceptual step.

AuthorsOpenReview ↗

Response to Reviewer woiC (2/5): Comparison Fairness and the Same-Backbone Control

Q2. Do the proposed method and the compared methods use the same source of object detections? If they use different detectors or localization procedures, how do the authors ensure a fair comparison?

No. AgentHOI takes its participant boxes from GroundingDINO, while HarnessHOI obtains them from the frozen MLLM in one full-image generation call. The detection sources therefore differ; fairness comes from three controls that fix the evaluator, then the backbone, then the downstream pipeline. All three configurations are evaluated under our common evaluator. Equation (11) defines class AP under the VOC 2007 eleven-point rule as AP_c = (1/11) Σ_t max_{r ≥ t} p_c(r) over the eleven recall thresholds, with mAP the mean of AP_c over the 600 classes (138 Rare, 462 Non-Rare). It depends only on the ranked predictions and the matching rule, so every row below is scored by the same rule. A common evaluator does not by itself equalize the inputs. For that we hold the backbone and then the downstream pipeline fixed in the two controls below. Appendix C.3 re-evaluates the released AgentHOI predictions with those matching and AP rules, and a third comparison holds the backbone fixed.

Configuration (common evaluator)Default FullDefault RareDefault Non-RareKO FullKO RareKO Non-Rare
AgentHOI, released GPT-4o + GroundingDINO predictions29.6840.4926.4533.7143.3930.82
AgentHOI, local Qwen3.5-9B + GroundingDINO-T28.2835.2926.1932.0739.7929.77
HarnessHOI, single frozen Qwen3.5-9B33.2041.6330.6836.8544.2034.65

AgentHOI's paper lists GPT-4o or Qwen2.5-VL-72B as its MLLM, with GroundingDINO for participant boxes. The released predictions use GPT-4o. Our single frozen 9B open model uses neither GroundingDINO nor GPT-4o or Qwen2.5-VL-72B as a second system. AgentHOI's own reported HICO Default Full is 29.61, and our re-evaluation of its released predictions gives 29.68, so the two evaluators agree within 0.07 points. Under that common evaluator, HarnessHOI reaches 33.20 Default Full and 36.85 Known Object Full, a 3.52 / 3.14 point margin over the 29.68 / 33.71 of the released predictions (Appendix C.3). On V-COCO, the strongest previous training-free entry reports 42.80 / 47.82 S1/S2 against our 44.65 / 49.70.

For the local adaptation we kept AgentHOI's three-stage prompts, scoring and post-processing, and changed only what local decoding required: output format, description length and the computation of generation-text probabilities. Both sides of this comparison therefore run the same Qwen3.5-9B, so the 4.92 Default Full and 4.78 KO Full point gaps isolate the pipeline rather than the backbone. AgentHOI reports 29.61 with GPT-4o and 26.12 with Qwen2.5-VL-72B, both with GroundingDINO. Our re-evaluation of the released GPT-4o predictions gives 29.68, while our 33.20 comes from one frozen 9B open model. Appendix D.6 of the submitted paper shows the same harness pattern on InternVL3-8B (32.11 / 43.60 / 48.40) and Qwen2.5-VL-7B (31.61 / 43.10 / 47.90), against 33.20 / 44.65 / 49.70 on Qwen3.5-9B. The comparison therefore holds both with the backbone fixed and under the released large-backbone configuration.

A second control changes only where the initial instances come from inside our own pipeline, with every downstream stage fixed. Appendix D.7 reports the localizer swap in full. The same-backbone control gives 28.28 against 33.20 HICO Default Full, and the localizer swap in Appendix D.7 gives 32.81 against 29.06 for the same template and downstream stages. The gap persists in each of the three controls.


AuthorsOpenReview ↗

Response to Reviewer woiC (3/5): Isolating the Harness from Initial Localization

W2. The improvements from individual components are relatively modest. The paper reports an increase in HICO-DET mAP from 31.45 to 33.20 and improvements on V-COCO from 42.90 to 44.65 and from 47.20 to 49.70. More detailed analysis would help establish which components contribute most to the overall performance.

Q3. Can the authors provide controlled experiments that isolate the contribution of the proposed perception and reasoning framework from the effects of the initial object localization process?

Yes. Tables 3 and 4 share the initial localization. Table 4 matches discovery and observation budgets, or fixes candidates and views for geometry and verification. We added paired three-seed runs of the same design, reported in Table 5 of the revised PDF. The initial-instance baseline and every component control share identical initial persons, objects and boxes. The baseline already performs visual pair assessment and relation verification over those participants. Discovery adds participants and observation supplies local evidence and grounding updates from the common initialization. These stages and geometric adjudication raise HICO Default Full from 31.45 to 33.20 and V-COCO S1/S2 from 42.90/47.20 to 44.65/49.70. Repeating this as a paired three-seed experiment gives +1.72 ± 0.21 HICO Default Full, +1.84 ± 0.17 HICO Known Object Full and +1.71 ± 0.11 / +2.46 ± 0.10 V-COCO S1/S2.

Table 4 orders the controls by HICO Default Full: 32.45 repeated localization, 32.50 relation proposal at verification, 32.55 excluding discovery-added participants, 32.70 non-directed observation, 32.79 relation proposal at box selection, 32.80 local observation disabled, 32.95 first candidate box only, against 33.20 for the full method.

Configuration (same Qwen3.5-9B, same initial boxes)HICO Default FullV-COCO S1V-COCO S2
HarnessHOI33.2044.6549.70
Initial-instance baseline31.4542.9047.20
Excluding discovery-added participants32.5544.0248.93
Local observation disabled32.8044.3349.21
First candidate box only32.9544.3849.48
Relation proposal at box selection32.7944.1849.16
Relation proposal at verification32.5043.9248.84
Repeated participant localization32.4543.8748.88
Non-directed observation32.7044.0749.08

Replacing discovery with repeated localization costs 0.75 HICO points, and non-directed observation at the same count and region size costs 0.50; both keep the budget fixed.

We also asked whether a better ranking of baseline candidates could explain the gap, fixing the predicted boxes, labels and candidate counts and ranking them by ground truth under the evaluator's matching rule.

Candidate and scoring conditionResponsesHICO Default FullV-COCO S1V-COCO S2
Direct generation, GT ranking121.5830.6935.01
Eight merged generations, exact-match, GT ranking830.5041.9046.40
Eight merged generations, pair-IoU, GT ranking830.0941.4045.90
HarnessHOI, predicted scores—33.2044.6549.70

With predicted scores, direct generation gives 15.62 / 16.41 / 21.20 and eight merged generations with pair-IoU suppression give 19.84 / 22.21 / 26.50. The table above instead ranks the same fixed candidates by ground truth: one response reaches 21.58, eight reach 30.50 with exact-match fusion or 30.09 with pair-IoU suppression. Even that oracle ranking stays below our 33.20. The retained pair set K = P_D ∪ P_O ∪ {p ∈ P : q_F(p) ≤ max(q_T(p), q_U(p))} takes the discovery and observation associations as inputs, and the scored domain Q_+ contains exactly the triplets whose pair lies in K. Discovery and observation change the candidate domain and the evidence used by s(h,a,o) = q_theta(true | x_rel) in Equation (6). The component comparisons share initial localization; GT ranking fixes each baseline's generated boxes, labels and candidates. Together, they show that interaction-guided perception improves candidate construction and supplies evidence for grounded relation judgments.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers

We sincerely thank all three reviewers for their detailed and constructive feedback. We appreciate their recognition of the interaction-centric integration of perception and reasoning (woiC, nNkX), the strong training-free results achieved with a single frozen MLLM without auxiliary grounding models (nNkX, W6rY), and the controlled comparisons supporting the proposed design choices (nNkX, W6rY). In particular, Reviewer W6rY highlights the explicit separation of participant grounding and interaction verification as a strength, while Reviewer nNkX describes the framework as “intuitive and rational, with clearly defined roles for each component”. We also appreciate the positive assessments of presentation and soundness from Reviewers woiC and nNkX.

We have revised the manuscript and added targeted experiments to strengthen the evidence for these contributions. The main additions are summarized below.

  1. Controlled comparisons with AgentHOI and alternative localization. Existing HOI detectors already use different backbones and pretraining strategies, and we compare HarnessHOI with them under the same HICO-DET protocol for end-to-end detection. When both systems use Qwen3.5-9B for reasoning, HarnessHOI achieves 33.20 mAP on HICO-DET Default Full compared with 28.28 for our local adaptation of AgentHOI. A separate shared-description control changes the initial localizer while retaining the downstream algorithm (Appendices C.3 and D.7).

  2. Direct perceptual diagnostics and component reliability. On both complete test sets, we measure observation-region accuracy, localization correction and unique-participant recovery, and evaluate GT visual inputs and observation-trigger sensitivity. Paired three-seed ablations give a +1.72 ± 0.21 HICO Default Full gain over the initial-instance baseline (mean ± sample SD; Tables 5–6 and Appendices D.3, D.8 and E.5).

  3. Transfer across benchmarks and backbones. Without benchmark-specific retuning, the Qwen3.5-9B harness reaches 12.62 Full AP on SWiG-HOI under THID, compared with 9.41 for the same-model initial-instance baseline. The existing HICO-DET/V-COCO results on InternVL3-8B and Qwen2.5-VL-7B complement this new benchmark evaluation (Appendices C.7 and D.6).

  4. Offline supervision for a downstream detector. Following the teacher-student suggestion, we use HarnessHOI to generate pseudo-labels for UPT-R50. With positive and non-positive label counts matched per image, the student gains +3.13 ± 0.17 HICO Default Full over supervision from the initial-instance teacher (Appendix H).

  5. Development and inference details. We specify that development annotation feedback excludes the union of Unseen classes across all four open-vocabulary partitions and is restricted to 254 common Seen classes. We also provide the full localization prompt and output format, model-request counts and stage costs, distinguishing GPU-worker computation from wall-clock latency (Appendices B.2, G.1 and E.2).

The revised PDF highlights the additions in blue, including the new component and perceptual analyses in Tables 5 and 6 of the main text and the supporting results in the appendices. We appreciate the reviewers’ feedback throughout the discussion, which helped us clarify the innovation with prior methods and explain the gains of HarnessHOI.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.