acceptodds
Under review as a conference paper at ICLR 2027

OQ-Atoms: Deferred Semantic Commitment for Open-Vocabulary 3D Segmentation

Abstract

Open-vocabulary 3D segmentation is fundamental to indoor embodied agents, whose semantic maps are built incrementally as RGB-D observations stream in. However, existing systems collapse multi-view semantic alternatives before the query is known, while learned refiners often over-commit and degrade reliable predictions. We present OQ-Atoms (Open-vocabulary Query-addressable Atoms), a three-stage framework that recasts semantic map updating as a query-time commitment decision. OQ-Atoms groups dense voxel evidence into compact atoms that preserve per-view observation provenance until query time, enabling class scores to be read out against original multi-view support rather than an averaged posterior. A conservative committer accepts an update only when four label-free evidence gates pass, keeping no-update as the default action. With fixed atoms and no per-scene tuning, OQ-Atoms reaches 41.63 mIoU on Replica and 40.13 mIoU on ScanNet. The accuracy gain comes from provenance-weighted readout, while the committer is evaluated separately as a safety rule. Atom construction costs two orders of magnitude less than semantic scoring. We further introduce leave-one-scene-out calibration for label-free semantic committers, showing that same-set gains can be threshold artifacts and that deployment rules should be judged by held-out abstention behavior.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
open-vocabulary 3D segmentation, semantic mapping, query-addressable representation, selective prediction, uncertainty-aware map updating

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DpiqOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes OQ-ATOMS for open-vocabulary 3D semantic segmentation. The method introduces query-addressable atoms that preserve multi-view semantic evidence, together with provenance-weighted readout and a conservative evidence-gated update mechanism. Experiments on Replica and ScanNet report improvements over existing methods. However, most of the reported performance advantage comes from the underlying dense backbone rather than the proposed components, while the deferred commitment mechanism is evaluated separately from the main results. Furthermore, inconsistencies and duplicated baseline visualizations in Figure 4 raise concerns about the reliability of the experimental presentation.

Critical strengths

None.

Critical weaknesses
  1. Serious inconsistencies and insufficiently supported experimental results.

Figure 4 raises substantial concerns about the reliability of the experimental presentation. The visualizations for HOV-SG and RayFronts are identical in both Replica examples, despite their substantially different quantitative results in Table 1. The figure caption also incorrectly identifies the third column as the proposed dense evidence stage. Moreover, the qualitative results provide little visible evidence of the substantial improvements claimed in Table 1. These inconsistencies make it difficult to assess whether the reported visualizations faithfully represent the evaluated methods.

  1. Unclear technical contribution.

The main performance gains appear to come primarily from the dense backbone rather than the proposed method. According to Table 2, the actual improvements from atoms and provenance readout are only 2.08 and 0.49 mIoU points on Replica and ScanNet, respectively. Meanwhile, the deferred commitment mechanism emphasized throughout the paper is not responsible for the headline results.

Questions for authors

None.

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

This paper proposes OQ-Atoms, which preserves multi-view evidence through atom grouping and provenance-weighted readout for open-vocabulary 3D segmentation. The paper also investigates semantic commitment through a separate rule-based mechanism, evaluated primarily in diagnostic experiments.

Critical strengths

The motivation is reasonable, and the method achieves performance improvements on Replica and ScanNet with clear ablation studies.

Critical weaknesses
  1. Limited novelty and performance gains. The core design relies on relatively simple spatial grouping and multi-view evidence aggregation. Moreover, the substantial gains in Table 1 largely come from the strong dense backbone, while the proposed modules provide only modest improvements.
  2. Limited effectiveness of semantic commitment. While the paper identifies over-commitment as a meaningful problem, the proposed heuristic mechanism does not demonstrate consistent benefits across held-out scenes. Its practical value beyond conservative abstention remains unclear.
Questions for authors
  1. Could the authors provide a Voxel + Provenance baseline and comparisons with simpler spatial grouping methods to better justify the necessity of atom grouping?
  2. Given the limited improvements under LOSO calibration, could the authors clarify what practical benefits the commitment mechanism provides beyond simply abstaining from updates?
Reviewer_2ZmVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents OQ-ATOMS for open-vocabulary 3D semantic segmentation. It groups voxels into atoms and preserves per-view evidence for provenance-weighted readout. Separate diagnostics examine whether proposed semantic corrections should be accepted through four evidence gates. Preserving observation evidence is a useful idea, but the main prediction path is insufficiently specified, and the evaluation leaves concerns about efficiency and competitiveness. These limitations outweigh the reported gains.

Critical strengths
  • Table 2 reports gains of 2.08 and 0.49 mIoU percentage points over the dense backbone on Replica and ScanNet, respectively. This provides evidence for the combined benefit of atom grouping and provenance-weighted readout.
  • The paper distinguishes segmentation accuracy from update safety and examines gate-threshold transfer through leave-one-scene-out calibration. Its explicit acknowledgment that the gates do not guarantee non-regression is useful, although the analysis does not establish safety of the full pipeline on unseen scenes.
Critical weaknesses
  • The main prediction path is insufficiently specified. Although atom construction and the CAQD architecture are described, the model generating per-view semantic scores and their confidence-weighted fusion into (S_0) are unclear. The final provenance-weighted readout is also not fully defined, including how stored score vectors support new queries. These omissions make the main results difficult to reproduce.
  • Practical efficiency and baseline coverage are insufficiently evaluated. The paper reports construction and semantic-scoring timings, but not the memory footprint or additional readout cost of retaining per-view evidence. Relevant methods such as Mosaic3D [1], OV3D [2], and RegionPLC [3] are also absent, without explaining whether differences in training or evaluation settings preclude comparison.

[1] Lee et al., "Mosaic3d: Foundation dataset and model for open-vocabulary 3d segmentation", CVPR, 2025. [2] Jiang et al., "Open-Vocabulary 3D Semantic Segmentation with Foundation Models", CVPR, 2024. [3] Yang et al., "RegionPLC: Regional Point-Language Contrastive Learning for Open-World 3D Scene Understanding", CVPR, 2024.

Questions for authors
  • Which model and checkpoint generate the per-view semantic scores, and how are these fused into S0S_0? Please also specify the final provenance-weighted readout and explain how stored score vectors support new queries.
  • Beyond the timings in Sec. 4.8, what additional memory and query latency does provenance-weighted readout require relative to the dense backbone?
  • Do differences in training or evaluation settings prevent comparison with Mosaic3D, OV3D, and RegionPLC? If not, are results under matched evaluation conditions available?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.