acceptodds
Under review as a conference paper at ICLR 2027

ConfluxCount: Forming and Enumerating Semantic Confluences for Text-Guided Object Counting

Abstract

Text-guided object counting requires converting spatially distributed query evidence into one count per object, while retaining weak responses and avoiding duplicate counts. Existing density and detection readouts respectively preserve coverage or expose instances, but neither explicitly organizes distributed evidence into object-level counting units before enumeration. We introduce ConfluxCount, which instead counts semantic confluences: object-centered structures formed by a text-grounded semantic flow over a dense probe lattice. A crowding-adaptive determinantal point process (DPP) then enumerates these confluences, suppressing redundant probes while preserving nearby instances, and produces both counts and locations from a unified probabilistic readout. On CLOC-v1.1, the jointly trained model reduces MAE from 8.39 for Count Anything to 6.03. It also reaches 4.14 MAE on KubriCount with category text and 5.75 on MixCount with positive text in supporting evaluations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Open-Vocabulary Counting, Text-Guided Counting, Semantic Confluence, Visual Grounding

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mQ66OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses text-guided object counting by transporting dense probes toward object-centered semantic confluences, followed by crowding-adaptive DPP enumeration. The method combines SAM3-based grounding, learned probe dynamics, and point-supervised training. Experiments are conducted on CLOC-v1.1, KubriCount, and MixCount to evaluate the proposed method.

Critical strengths
  1. The motivation is reasonable and clear. Separating evidence grouping from enumeration provides a coherent approach to duplicate suppression, and DPP offers a mechanism for balancing competition and coexistence among nearby probes.
  2. The proposed method achieves outstanding counting performance.
  3. The authors provide diagnostic analyses, including per-object inclusion mass, probe contraction, explicit dense-scene failure analysis, and informative investigation of probe-resolution limitations.
Critical weaknesses
  1. Some results contradict the metric definition. For instance, the values of NAE and SNAE(2) of ConfluxCount on NWPU-VHR-10 in tab. 10 seem to violate SNAE(2)≥NAE\mathrm{SNAE}(2)\geq\mathrm{NAE} according to Appendix A.1. The authors are recommended to verify these values and clarify whether global and spatial metrics are computed consistently.
  2. The pure-shared DPP ablation appears inconsistent with the stated parameterization. Under the 64-rank setting in Appendix A.3, setting π=0\pi=0 yields L=UU⊤L=UU^\top, implying tr⁡(K)≤rank⁡(L)≤64\operatorname{tr}(K)\leq\operatorname{rank}(L)\leq64. Using the count-bin minima and test-sample frequencies in fig. 8, the conservative full-test MAE lower bound is approximately 12.26, exceeding 6.89 reported in tab. 6. Does this ablation retain a nonzero diagonal component, change the factor rank, or use a different evaluation protocol? Otherwise, the reported result is incompatible with the stated formulation.
  3. The native DPP localization readout shows limited practical advantage. With the model and predicted count fixed, tab. 7 shows that greedy DPP underperforms marginal-ranked NMS in F1, R@8, and GAME1, matching the localization error. This may limit the demonstrated benefit of the proposed localization decoder.
  4. The training protocol underlying the headline results is insufficiently documented. (1) The main tables use joint training, whereas Appendix A details the CLOC-only recipe. The authors are encouraged to provide exact training splits, dataset-mixing ratios, and the joint optimization schedule. Although the CLOC-only results provide a useful control, comparisons on matched training data would better isolate methodological gains on the supporting benchmarks. (2) Appendix B.3 attributes mismatch robustness partly to empty-set supervision, but the construction, sampling frequency, and loss treatment of mismatched training pairs are unspecified; the in-batch contrastive objective alone does not establish this supervision.
  5. Canonical-probe coverage introduces an unquantified supervision bottleneck. Appendix A.3 allows GT objects with empty probe Voronoi cells to lack a canonical representative for both the DPP and coverage losses. Consequently, the exact-set likelihood may favor a target of size ∣S⋆∣<N|S^\star|<N, while the count loss still targets the full count NN. This potential supervision mismatch may contribute to dense-scene undercounting. Reporting the uncovered GT fraction 1−∣S⋆∣/N1-|S^\star|/N across local densities and lattice resolutions, alongside density-stratified per-object inclusion mass, would help distinguish supervision limitations from inference capacity.
  6. The paper involves numerous components, including semantic flow, instance memory, basin pyramid field, probe dynamics, a crowding gate, DPP, and multiple loss functions. The main text is dense with exposition and notation, and is difficult for readers to quickly identify which components or designs are truly essential.
Questions for authors

Please refer to the weaknesses above.

Reviewer_DtmmOpenReview ↗
Rating
1
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies text-guided object counting and localization. Existing counting paradigms face complementary limitations: density-based methods lack explicit object identity, while detection-style methods expose discrete candidates yet remain vulnerable in crowded scenes. The proposed ConfluxCount addresses this by moving a dense probe lattice through a query-conditioned semantic flow so that evidence from the same object forms an object-centered "semantic confluence", followed by a crowding-adaptive determinantal point process (DPP) that assigns cardinality and locations from the resulting probe configuration. The jointly trained model reports SOTA across several benchmarks.

Critical strengths
  1. Meaningful intermediate representation. Separating object formation from cardinality assignment bridges dense evidence and instance-level counting.

  2. Strong benchmark results. The method delivers strong performance and consistent improvements across public benchmarks.

  3. Honest analysis of failure modes. The paper provides comprehensive ablations and detailed analyses of bad cases.

Critical weaknesses

Major Weaknesses

  1. Confluence gains are small and attractors are insufficiently supported. The ablations suggest that semantic confluence has limited impact on performance. Figures 4 and 5 rely mainly on GT-defined basins and do not establish unique attractors or their split/merge behavior. The paper also lacks convincing evidence, such as attractor visualizations, that the learned attractors perform the claimed role.

  2. ConfluxCount is not convincingly specific to text-guided counting. For motivation, treating thresholding as a fundamental defect of point-based counting is not well motivated, as far as I known. More importantly, the paper reads largely as a general counting framework, while the ablations provide limited evidence that 'semantic confluence' captures or benefits from text-specific information.

  3. DPP lacks instance-level guarantees. Its localization metrics trail NMS, and quality/diversity does not imply physical instance identity. Moreover, canonical probe routing may leave some ground-truth instances without a representative, resulting in insufficient supervision coverage.

  4. Supervision are asymmetric inconsistent. The main results use a model jointly trained on three datasets, while the baselines do not necessarily use matched training or prompt conditions. In addition, Sections 3.5 and B.3 give conflicting accounts of the mismatch-query and empty-set supervision.

Minor issues

  1. The predefined probe budget imposes a capacity and efficiency trade-off. A fixed number of probes cannot handle scenes whose object count exceeds the available capacity, while increasing this bound may substantially affect computational efficiency.

  2. SNAE definition and reported values need reconciliation. Appendix A.1 defines SNAE(0) = NAE and claims monotonicity with partition level, but Table 10 reports NAE = 1.328 and SNAE(2) = 1.194 for NWPU-VHR-10. This may reflect an implementation, aggregation, or table error and should be explained with raw results.

Questions for authors
  1. Confluence and attractors. Beyond GT-defined basin statistics, what evidence shows that individual objects form distinct attractors without splitting or merging?

  2. Impact of text-conditioned confluence. Can text-conditioned confluence contribute more substantially to overall performance within the proposed framework?

  3. The role of DPP. The paper places considerable emphasis on DPP, which yields substantial performance gains. However, these gains seem largely unrelated to text conditioning (closer to generic counting).

Reviewer_LN1bOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ConfluxCount, a text-guided object counting framework that combines semantic flow-based probe aggregation with crowding-adaptive Determinantal Point Process (DPP) enumeration. The method organizes dense visual evidence into object-centered semantic confluences and estimates object counts through the expected DPP cardinality. Experiments on CLOC-v1.1, KubriCount, and MixCount report competitive counting performance.

Critical strengths
  1. Competitive counting performance. The method achieves strong results across multiple benchmarks, particularly on CLOC-v1.1, where it improves MAE from 8.39 to 6.03 compared with Count Anything.

  2. Reasonable methodological motivation. Combining semantic evidence aggregation with adaptive redundancy suppression provides an intuitive approach to handling duplicated responses and crowded objects.

Critical weaknesses
  1. The improvements attributable to the proposed components are relatively modest. In Table 4, DPP achieves 6.12 MAE compared with 6.55 for density integration, while Table 8 shows only a small improvement from adding probe transitions (6.30 to 6.12).

  2. The readout comparisons use representations optimized for the proposed framework, rather than independently trained alternatives, making it difficult to establish the actual advantage of DPP enumeration.

  3. Table 7 further shows that the DPP-based localization strategy underperforms marginal-based NMS on F1 and recall, weakening the claimed benefit of the unified counting and localization framework.

  4. The main results use joint training across three datasets, whereas competing methods may use different training data and prompting protocols. This makes the overall performance improvements difficult to attribute solely to the proposed methodology.

  5. The paper lacks sufficiently strong end-to-end comparisons with simpler alternatives under identical backbone, training data, and optimization settings.

  6. The joint-training details are not fully documented, limiting the reproducibility and interpretation of the reported improvements.

Questions for authors
  1. How much of the performance gain comes specifically from semantic flow and DPP enumeration? Could the authors provide end-to-end comparisons against density-based counting and learned suppression methods under identical training conditions?

  2. Why are the improvements from the core components relatively small? In particular, how do the authors justify the additional complexity of probe transitions and DPP enumeration given the results in Tables 4 and 8?

  3. Can the authors clarify the joint-training protocol and its influence on performance? How would the method compare with the strongest baselines when using exactly the same training data and prompting interfaces?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.