acceptodds
Under review as a conference paper at ICLR 2027

EviSplat: Evidence-Preserving Language Fields for 3D Gaussian Splatting

Abstract

Open-vocabulary 3D scene understanding enables object localization and segmentation from free-form text queries without a fixed category vocabulary. Many recent methods build on 3D Gaussian Splatting and consolidate multi-view observations, such as masked crops from individual views, into language features or compact object descriptors before the query is known. However, observations of the same object vary across viewpoints and are not equally informative: some reveal cues relevant to a particular query, whereas others provide incomplete or misleading evidence. Pre-query consolidation can therefore suppress cues on which a later query depends. We introduce EviSplat, which preserves individual observation features as evidence for later text queries. EviSplat retains individual observation features within class-agnostic 3D instances that represent objects, object parts, or background regions. It also learns, for each Gaussian, a distribution describing which visual appearances its observations support. Given a text query, EviSplat scores each instance using its most relevant observations. It then computes a score for each Gaussian by combining instance-level relevance with locally supported evidence, weighted by how often and how unambiguously that Gaussian was observed. Different queries can thus draw on different visual cues from the same preserved evidence. Experiments across diverse datasets and evaluation protocols demonstrate state-of-the-art performance, supporting the benefit of preserving multi-view evidence until query time and aggregating it according to the query.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Open-vocabulary 3D segmentation, 3D Gaussian Splatting

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_uBFHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

EviSplat combines instance observation banks and a Gaussian evidence field to preserve multi-view evidence for query-guided open-vocabulary 3D segmentation. Experiments show gains over a reproduced LaGa baseline, with benefits from observation selection and Gaussian-local evidence. However, the novelty is relatively weak compared with prior work, while the learned evidence field and reliability weighting offer only modest gains over sparse empirical distributions and unweighted fusion.

Critical strengths
  1. The controlled ablations isolate the contribution of top-β observation selection through comparisons with mean and max pooling.
  2. The component analysis separately evaluates Gaussian-local evidence and fixed-mass aggregation, helping distinguish the contributions of local evidence and aggregation design.
  3. EviSplat is less sensitive to source-feature corruption.
  4. The D=16 Gaussian evidence field matches sparse empirical distributions in accuracy while reducing local representation storage.
Critical weaknesses
  1. Instance observation banks and top-β observation selection overlap with the per-view CLIP embeddings and top-k relevance averaging of https://arxiv.org/abs/2509.12938. Their distinct contribution remains insufficiently isolated.
  2. Most baseline results are cited as reported. The claimed 5.1-point lead involves different query protocols and lacks matched re-evaluation; the controlled LaGa comparison provides stronger evidence.
  3. The learned Gaussian evidence field offers limited accuracy gains over sparse empirical distributions despite higher local storage and query costs.
  4. Observation reliability provides limited additional gains: removing or modifying reliability weighting has little effect on performance in the reported single-run ablations.
  5. Gains are uneven across datasets and scenes, with marginal improvement on 3D-OVS and degraded performance on RAMEN after adding Gaussian-local evidence. These failure cases remain insufficiently explained.
Questions for authors
  1. How does https://arxiv.org/abs/2509.12938’s top-k relevance averaging compare with top-β observation selection under the same LaGa backbone, masks, CLIP encoder, grouping, and post-processing?
  2. What are the total storage and end-to-end query latency for D=16, D=32, and sparse empirical distributions?
  3. Given that β was selected on LERF-OVS, how were γ and T selected? Are the results robust when these parameters are chosen using held-out scenes or datasets?
  4. Would the strongest baselines in Tables 1 and C.1 retain their reported rankings under matched backbones, query granularity, post-processing, and averaging conventions?
2 comments

Thank the authors for the additional experiments, which address my main concerns about the controlled comparisons, hyperparameter selection, and computational costs. I still have concerns about novelty, as prior work already retains multi-view features and combines their relevance scores for each query. Nevertheless, this work provides a marked improvement, and I appreciate the empirical value of this simple yet effective extension. On balance, given the convincing experimental evidence provided in the rebuttal, I tend to raise my rating to 3 (Weak acceptance), while retaining my reservations about the technical novelty.

We sincerely thank the reviewer for the thoughtful follow-up and for raising the rating. It has been a privilege to strengthen our paper through your insightful feedback, and we are very glad that the additional experiments addressed your main concerns. Please feel free to reach out if there are any further points you would like to discuss. We wish you a wonderful day!

Reviewer_VQ5ZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EviSplat is a query-time scoring layer on top of LaGa for open-vocabulary 3D segmentation in 3DGS. It keeps all per-view CLIP features per instance (since observations of the same object vary across viewpoints and are not equally informative: some reveal cues relevant to a query, whereas others provide incomplete or misleading evidence) and scores by the top-k mean of their relevance, and adds a per-Gaussian prototype histogram learned as a compact latent whose relevance is aggregated over a fixed probability mass and fused with the instance score. On LERF-OVS it reaches 70.7 mIoU vs. 63.6 for reproduced LaGa and 65.6 for the best prior method in 2D rendered-feature querying.

Critical strengths

A simple yet reasonable idea with a large, well-ablated gain. Scoring observations individually and selecting at query time is well motivated and gives +7.1 mIoU over the reproduced baseline and +5.1 over the best prior method on LERF-OVS. The ablations are thorough and honest about which components do not help.

Critical weaknesses
  1. The instance-level contribution is prior work, and the Gaussian-level machinery adds little. The instance prior is the bag-of-embeddings scoring of Arafa & Stricker (2025), cited in one sentence but never compared against. LaGa itself already keeps up to 20 query-reweighted descriptors per object, so the "pre-query consolidation discards cues" framing is overstated.
  2. Hyperparameters are tuned on the four-scene test set, and cost is under-reported. End-to-end query time and total storage versus LaGa are not reported.
Questions for authors
  1. Please compare with Arafa & Stricker (2025).
  2. Can you validate (β, γ, T) and the post-processing thresholds on held-out data?
  3. What are the end-to-end query time and total storage versus LaGa?
Reviewer_vavUOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

EviSplat advances open-vocabulary 3D Gaussian Splatting segmentation by addressing the information loss caused by compressing per-view CLIP features before a query is known. It pre-computes two representations: instance observation banks that preserve raw per-view features within class-agnostic clusters, and a Gaussian evidence field that encodes local feature distributions into scene-specific codebook histograms. At query time, EviSplat fuses top instance-level LERF relevance with high-confidence local evidence via a positive-only correction, producing final masks through LaGa’s thresholding and rendering while outperforming baseline methods across benchmark datasets.

Critical strengths
  • EviSplat’s primary strength lies in its simple yet useful training-free component. By scoring individual retained observations and utilizing a top-beta mean instead of a consolidated descriptor, the approach yields performance gains without requiring extra training. This insight offers a practical modification that other frameworks can adopt.

  • Additionally, the authors provide thorough evaluations, including matched-checkpoint ablations, repeated runs, and rigorous storage and timing measurements, while openly disclosing unfavorable outcomes like memory overhead and benchmark-selected hyperparameters. This transparency makes the findings auditable and establishes credibility for the work.

Critical weaknesses

The pipeline cannot return an empty result, so it is not evaluated as a true free-form query system

  • The system always produces a mask, even for missing objects. After multi-scale averaging, scores are min–max normalized and thresholded at 0.7, ensuring that the highest-scoring Gaussians receive a score of 1.0. However, this relative thresholding provides no explicit mechanism for rejecting absent objects, while the positive-only fusion rule can only increase scores. Moreover, in multi-target queries, min–max normalization can suppress weaker but valid matches when a single instance dominates the score distribution, potentially leading to false negatives.
Questions for authors
  • Robustness to absent and multiple targets. Can the authors evaluate absent-target queries using labels from other LERF-OVS scenes, reporting false-positive rates or predicted mask areas for the baseline, instance-only variant, and full model? Additionally, for queries with multiple disjoint ground-truth instances (e.g., ``knife''), does relative thresholding suppress weaker secondary matches?

  • Dependence on thresholding and the LaGa cosine gate. How much of the improvement depends on min--max normalization, quantile clipping, and the LaGa cosine-similarity gate? Could the authors compare against a fixed absolute threshold on fused relevance scores (e.g., 0.5) and report whether the gains over the baseline persist?

  • Generalization and sensitivity to hyperparameter selection. Since β\beta, γ\gamma, and TT were selected on the four LERF-OVS test scenes, could the authors report results under leave-one-scene-out hyperparameter selection? This would help establish whether the reported gains generalize beyond the scenes used for tuning.

2 comments

Thanks for the authors detailed comments, they have addressed my concerns and I will keep my score.

We sincerely thank the reviewer for the follow-up and are glad that our responses addressed your concerns. Your feedback has helped us improve the paper, and we will reflect the discussion in the revision. We would be happy to discuss any further points at any time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.