acceptodds
Under review as a conference paper at ICLR 2027

GaussianCollab: Communication-Efficient Latent Query Fusion via Tokenized Semantic Gaussians for Collaborative Occupancy Prediction

Abstract

Collaborative perception requires agents to exchange spatially grounded semantic evidence under constrained communication budgets. Existing approaches predominantly communicate dense bird's-eye-view or voxel features, which either collapse vertical geometry or incur the cost of dense 3D grids. We propose , a collaborative perception framework that uses tokenized semantic Gaussians as the inter-agent communication medium. GaussianCollab communicates and fuses latent semantic queries anchored to explicit Gaussian geometry. This shifts Gaussian-based collaboration from attribute-only exchange to latent-query fusion, preserving explicit 3D structure for rigid transformation and geometric matching while retaining richer learned semantics. GaussianCollab performs geometry-guided matching and latent-query fusion over matched tokens, and decodes and stacks unmatched collaborator evidence. To control the resulting payload, we introduce together with . Experiments on Semantic-OPV2V and Co3SOP demonstrate state-of-the-art performance on collaborative semantic occupancy prediction task. Under matched Gaussian capacity, the lowest-communication-volume operating points improve mIoU over the reproduced VOGS-L baseline by while reducing communication volume by % and % on Semantic-OPV2V and Co3SOP, respectively. GaussianCollab also maintains favorable accuracy under the tested token-erasure and localization-perturbation settings. These results establish geometry-anchored latent Gaussian queries as an effective and communication-efficient fusion medium for vision-based collaborative occupancy prediction. Code will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
collaborative perception, semantic occupancy prediction, semantic Gaussians, 3D Gaussian Splatting, latent fusion

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_Jsx7OpenReview ↗
Rating
4(+1)
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a framework for 3D semantic occupancy prediction, GaussianCollab. By transmitting tokenized representation, the proposed method achieved better overall results on the tested Semantic-OPV2V and Co3SOP.

Critical strengths
  1. The communication medium and the corresponding collaboration protocol are well motivated, yielding an overall better result.
  2. The SIC-guided selection further improves the method from a communication-efficiency perspective.
Critical weaknesses

The benefit of using latent queries over decoded attributes is not fully justified. From GSCollab-112 N/I/T results we can see a ~2 mIoU degradation to achieve 5.5 -> 0.9 CV reduction. From the VOGS-L results, N Protocol decoded attributes can achieve similar CV as T protocol latent queries. This raises the question of whether a GSCollab N using decoded attributes can achieve similar performance. Further design justification or an additional comparison on a controlled CV will help address this concern.

Questions for authors

Please see the weaknesses section.

7 comments
AuthorsOpenReview ↗

Official Comment 1: Corrected CV and updated main results

We first provide corrected communication-volume (CV) results for PEQC-compressed GSCollab in the manuscript's table layout for easy comparison. We address the reviewers' questions in subsequent official comments. PEQC CV previously counted 128-dimensional queries reconstructed after reception rather than the compressed queries transmitted. The code correction changes byte accounting, not the trained models or prediction logic. We reevaluated the reported settings on the full validation split using their existing checkpoints; pose and packed SIC-request bytes remain included. Independent evaluations can yield different IoU/mIoU, so each updated accuracy value is reported with the CV measured in the same evaluation. Accuracy changes relative to the submission come from reevaluation, not from the byte-count correction. Unchanged baselines retain their manuscript values. The central result remains: compressed SIC settings exceed VOGS-L in semantic mIoU at lower CV.

CV is MiB per collaborative frame; IoU/mIoU are percentages. Slash-separated cells use CV/IoU/mIoU. N: no SIC; I: SIC at inference only; T: SIC in training and inference. ✓: stacking; ✓→--: stacking only during training. GSCollab-128 omits PEQC.

Table 2. Semantic-OPV2V fusion and communication controls

SettingFused contentStackProtocolIoUmIoUCV
VOGS-LDecoded attributes--N71.5438.900.941
GSCollab-128Latent queries--N72.8839.715.113
GSCollab-128Latent queries✓N74.1142.835.186
GSCollab-112Latent queries✓N73.9243.194.828
GSCollab-112Latent queries✓I73.2242.470.914
GSCollab-112Latent queries✓T72.4742.030.773
GSCollab-96 backboneDecoded attributes--T71.8638.910.196
GSCollab-96Latent queries✓→--T72.2939.900.695

Table 5. Co3SOP operating points (CV/IoU/mIoU)

MethodStackNIT
VOGS-L--0.786/77.74/41.19----
GSCollab-128--4.903/77.31/42.56----
GSCollab-128✓4.460/76.18/43.930.929/76.53/43.100.909/77.17/43.44
GSCollab-64✓2.533/78.06/44.440.523/77.94/43.900.511/77.89/44.45
GSCollab-80✓3.164/77.81/45.840.508/77.65/45.350.504/78.04/44.86
GSCollab-96✓3.884/77.63/44.970.613/77.51/44.540.566/77.71/44.67
GSCollab-112✓4.421/76.87/46.050.723/77.33/45.390.668/78.01/44.05

Table 8. SIC versus entropy-confidence gating

WidthHeur. mIoUSIC mIoUGainHeur. CV (MiB)SIC CV (MiB)CV reduction
12840.31841.501+1.1831.5761.10529.9%
6440.71241.938+1.2260.7630.51732.3%
8040.26340.913+0.6501.0600.73330.9%
9640.22041.618+1.3981.0700.89016.9%
11240.97142.474+1.5031.3440.91432.0%

Table 9. Semantic-OPV2V operating points (CV/IoU/mIoU)

MethodStackNIT
VOGS-L--0.941/71.54/38.90----
GSCollab-128--5.113/72.88/39.71----
GSCollab-128✓5.186/74.11/42.831.105/73.15/41.500.942/73.47/41.92
GSCollab-64✓2.923/74.43/42.490.517/73.67/41.940.484/73.65/41.47
GSCollab-80✓3.397/73.37/41.930.733/72.51/40.910.591/72.44/41.68
GSCollab-96✓4.022/74.69/42.340.890/73.73/41.620.695/72.87/42.64
GSCollab-112✓4.828/73.92/43.190.914/73.22/42.470.773/72.47/42.03

Table 10. Semantic-OPV2V erasure sweep (CV/IoU/mIoU)

Erasure rateVOGS-LGSCollab-64TGSCollab-80TGSCollab-96TGSCollab-112T
0%0.941/71.54/38.900.484/73.65/41.470.591/72.44/41.680.695/72.87/42.640.773/72.47/42.03
5%0.894/71.56/38.780.463/73.64/41.350.564/72.45/41.520.662/72.84/42.470.736/72.45/41.89
10%0.847/71.53/38.710.440/73.59/41.220.535/72.41/41.690.629/72.82/42.360.699/72.44/41.86
20%0.753/71.54/38.710.396/73.52/40.790.480/72.36/41.150.563/72.76/42.100.625/72.38/41.51
30%0.659/71.51/38.490.351/73.40/40.610.425/72.30/40.980.498/72.65/41.750.552/72.28/41.07
50%0.471/71.40/38.150.262/73.08/39.860.315/72.02/40.560.366/72.36/40.890.405/72.03/40.65

Table 10, CoHFF feature-sparsification reference (unchanged)

Metric0%50%80%95%99%
CV (MiB)15.7647.8873.1570.7920.162
IoU50.4649.5649.5348.5248.02
mIoU34.1632.9732.7030.1329.48
AuthorsOpenReview ↗

Official Comment 2 for Jsx7-Q1 — Is latent communication justified at controlled CV?

Question. “The benefit of using latent queries over decoded attributes is not fully justified. From GSCollab-112 N/I/T results we can see a ~2 mIoU degradation to achieve 5.5 -> 0.9 CV reduction. From the VOGS-L results, N Protocol decoded attributes can achieve similar CV as T protocol latent queries. This raises the question of whether a GSCollab N using decoded attributes can achieve similar performance. Further design justification or an additional comparison on a controlled CV will help address this concern.”

Response: We corrected communication-volume (CV) accounting for compressed GSCollab variants and reran validation. The updated main results appear in our separate official comment, “Corrected CV and updated main results.”

Additional experiments are underway to address your proposed comparison. We will present their accuracy and measured CV in one ablation table in a single shared follow-up official comment.

AuthorsOpenReview ↗

Official Comment 5: Consolidated controls for latent and attribute collaboration

These Semantic-OPV2V controls use 12,800 initial Gaussians, 0.4 m matching, and all 1,980 validation frames. N uses no SIC, I uses SIC at inference only, and T uses SIC in training and inference. CV (MiB per collaborative frame) counts the unfiltered response for N, or the packed request and selected response for I/T. IoU/mIoU are percentages.

#Backbone settingTransmitted ContentToken LengthFused ContentStackIoUmIoUCV
0VOGS-L (N)Shared Refinement Head Decoded Attributes7 pose + 17 attributes = 24 (96 B)Shared Refinement Head Decoded Attributes--71.5438.900.941
1GSC-112 (N)Shared Refinement Head Decoded Attributes7 pose + 17 attributes = 24 (96 B)Shared Refinement Head Decoded Attributes✓73.2241.900.976
2GSC-112 (I)Pose + Latent Queries7 pose + 112 query = 119 (476 B)Latent Queries✓73.2242.470.914
3GSC-112 (T)Pose + Latent Queries7 pose + 112 query = 119 (476 B)Latent Queries✓72.4742.030.773
4GSC-96 (T)Shared Refinement Head Decoded Attributes7 pose + 17 attributes = 24 (96 B)Shared Refinement Head Decoded Attributes--71.8638.910.196
5GSC-96 (T)Shared Refinement Head Decoded Attributes7 pose + 17 attributes = 24 (96 B)Shared Refinement Head Decoded Attributes✓73.2941.190.206
6GSC-96 (T)Pose + Latent Queries7 pose + 96 query = 103 (412 B)Latent Queries--72.5140.860.754
7GSC-96 (T)Pose + Latent Queries7 pose + 96 query = 103 (412 B)Latent Queries✓72.8742.640.695
8GSC-96 (T)Pose + Latent Queries7 pose + 96 query = 103 (412 B)Frame-Invariant Attr. Decoder Decoded Attributes--71.7237.450.926
9GSC-96 (T)Pose + Latent Queries7 pose + 96 query = 103 (412 B)Frame-Invariant Attr. Decoder Decoded Attributes✓72.8739.880.785
10GSC-17 (T)Pose + Latent Queries7 pose + 17 query = 24 (96 B)Latent Queries✓72.3842.260.189

Attribute controls. As requested by reviewers 3VWs and bVkA, row 5 provides the decoded-attribute, SIC-T, stacking control: 41.19 mIoU at 0.206 MiB, versus 38.91 at 0.196 MiB for the separately trained no-stacking row 4. Reproduced VOGS-L (row 0) does not retain unmatched neighbor Gaussians. Row 1 further reaches 41.90 mIoU at 0.976 MiB using decoded attributes without SIC. Thus, strong gains over VOGS-L are possible without transmitting latent queries; they cannot all be attributed to latent communication.

Fusion pathway. Rows 6/8 and 7/9 transmit the same pose-plus-96-D-query format and use the same SIC and stacking settings within each pair. The direct-query-fusion settings achieve 3.41 and 2.76 more mIoU points, respectively, than decoding the queries into attributes before fusion, with lower measured CV in both comparisons. This supports the latent fusion pathway. Unlike row 5, rows 8–9 still transmit queries and decode them only after reception. The separately trained pairs therefore do not isolate the value of query transmission itself.

Communication budget. The 17-D latent token in row 10 has the same 24-float length as the attribute token in row 5. At similar total CV, it reaches 42.26 versus 41.19 mIoU (0.189 versus 0.206 MiB), while binary IoU is lower (72.38 versus 73.29). For Jsx7's proposed N-protocol attribute control, row 1 achieves mIoU close to the 112-D latent T setting in row 3 (41.90 versus 42.03); row 3 uses about 21% less CV, but also changes the SIC protocol. From row 2 to row 3, CV falls from 0.914 to 0.773 MiB while mIoU falls by 0.44 points. These results support the complete compact-latent design's semantic-mIoU/communication trade-off, rather than a universal advantage attributable to the representation alone.

Reviewer_3VWsOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GaussianCollab for communication-efficient multi-agent 3D semantic occupancy prediction. It transmits geometry-anchored latent queries, using semantic disagreement for token selection and attention-based fusion with unmatched Gaussian stacking. Experiments on Semantic-OPV2V and Co3SOP evaluate accuracy, communication costs, and component contributions.

Critical strengths
  1. The framework addresses a practical challenge in collaborative perception, combining geometry-anchored query fusion and unmatched-Gaussian stacking to integrate complementary observations under communication constraints.
  2. The full system improves mIoU by 2.95 points over VOGS-L at lower communication volume. Component ablations and robustness tests provide useful support for its effectiveness.
Critical weaknesses
  1. The latent-query interface’s independent efficiency advantage remains unclear. Table 2’s no-stacking comparison gains 1.03 mIoU points at 4.58 times the payload. A comparable-budget attribute baseline with SIC and stacking would better isolate its contribution.
  2. Cross-frame query consistency lacks direct validation. Separate pose and attribute encoders do not guarantee pose-independent queries, and existing single-agent ablations and pose-noise tests do not directly verify consistent recovery of frame-invariant attributes after coordinate transformation.
Questions for authors
  1. What gains do latent queries provide over explicit attributes when both use the same SIC selection and unmatched-Gaussian stacking mechanisms under comparable communication budgets?
  2. How is pose information handled in the final latent queries? Does re-expressing the same token in another coordinate frame preserve its decoded frame-invariant attributes and semantic predictions?
  3. Why are only the original scores max-normalized when computing the blending weight? Could the existing results also clarify whether SIC-selected disagreements lead to effective error correction?
7 comments
AuthorsOpenReview ↗

Official Comment 1: Corrected CV and updated main results

We first provide corrected communication-volume (CV) results for PEQC-compressed GSCollab in the manuscript's table layout for easy comparison. We address the reviewers' questions in subsequent official comments. PEQC CV previously counted 128-dimensional queries reconstructed after reception rather than the compressed queries transmitted. The code correction changes byte accounting, not the trained models or prediction logic. We reevaluated the reported settings on the full validation split using their existing checkpoints; pose and packed SIC-request bytes remain included. Independent evaluations can yield different IoU/mIoU, so each updated accuracy value is reported with the CV measured in the same evaluation. Accuracy changes relative to the submission come from reevaluation, not from the byte-count correction. Unchanged baselines retain their manuscript values. The central result remains: compressed SIC settings exceed VOGS-L in semantic mIoU at lower CV.

CV is MiB per collaborative frame; IoU/mIoU are percentages. Slash-separated cells use CV/IoU/mIoU. N: no SIC; I: SIC at inference only; T: SIC in training and inference. ✓: stacking; ✓→--: stacking only during training. GSCollab-128 omits PEQC.

Table 2. Semantic-OPV2V fusion and communication controls

SettingFused contentStackProtocolIoUmIoUCV
VOGS-LDecoded attributes--N71.5438.900.941
GSCollab-128Latent queries--N72.8839.715.113
GSCollab-128Latent queries✓N74.1142.835.186
GSCollab-112Latent queries✓N73.9243.194.828
GSCollab-112Latent queries✓I73.2242.470.914
GSCollab-112Latent queries✓T72.4742.030.773
GSCollab-96 backboneDecoded attributes--T71.8638.910.196
GSCollab-96Latent queries✓→--T72.2939.900.695

Table 5. Co3SOP operating points (CV/IoU/mIoU)

MethodStackNIT
VOGS-L--0.786/77.74/41.19----
GSCollab-128--4.903/77.31/42.56----
GSCollab-128✓4.460/76.18/43.930.929/76.53/43.100.909/77.17/43.44
GSCollab-64✓2.533/78.06/44.440.523/77.94/43.900.511/77.89/44.45
GSCollab-80✓3.164/77.81/45.840.508/77.65/45.350.504/78.04/44.86
GSCollab-96✓3.884/77.63/44.970.613/77.51/44.540.566/77.71/44.67
GSCollab-112✓4.421/76.87/46.050.723/77.33/45.390.668/78.01/44.05

Table 8. SIC versus entropy-confidence gating

WidthHeur. mIoUSIC mIoUGainHeur. CV (MiB)SIC CV (MiB)CV reduction
12840.31841.501+1.1831.5761.10529.9%
6440.71241.938+1.2260.7630.51732.3%
8040.26340.913+0.6501.0600.73330.9%
9640.22041.618+1.3981.0700.89016.9%
11240.97142.474+1.5031.3440.91432.0%

Table 9. Semantic-OPV2V operating points (CV/IoU/mIoU)

MethodStackNIT
VOGS-L--0.941/71.54/38.90----
GSCollab-128--5.113/72.88/39.71----
GSCollab-128✓5.186/74.11/42.831.105/73.15/41.500.942/73.47/41.92
GSCollab-64✓2.923/74.43/42.490.517/73.67/41.940.484/73.65/41.47
GSCollab-80✓3.397/73.37/41.930.733/72.51/40.910.591/72.44/41.68
GSCollab-96✓4.022/74.69/42.340.890/73.73/41.620.695/72.87/42.64
GSCollab-112✓4.828/73.92/43.190.914/73.22/42.470.773/72.47/42.03

Table 10. Semantic-OPV2V erasure sweep (CV/IoU/mIoU)

Erasure rateVOGS-LGSCollab-64TGSCollab-80TGSCollab-96TGSCollab-112T
0%0.941/71.54/38.900.484/73.65/41.470.591/72.44/41.680.695/72.87/42.640.773/72.47/42.03
5%0.894/71.56/38.780.463/73.64/41.350.564/72.45/41.520.662/72.84/42.470.736/72.45/41.89
10%0.847/71.53/38.710.440/73.59/41.220.535/72.41/41.690.629/72.82/42.360.699/72.44/41.86
20%0.753/71.54/38.710.396/73.52/40.790.480/72.36/41.150.563/72.76/42.100.625/72.38/41.51
30%0.659/71.51/38.490.351/73.40/40.610.425/72.30/40.980.498/72.65/41.750.552/72.28/41.07
50%0.471/71.40/38.150.262/73.08/39.860.315/72.02/40.560.366/72.36/40.890.405/72.03/40.65

Table 10, CoHFF feature-sparsification reference (unchanged)

Metric0%50%80%95%99%
CV (MiB)15.7647.8873.1570.7920.162
IoU50.4649.5649.5348.5248.02
mIoU34.1632.9732.7030.1329.48
AuthorsOpenReview ↗

Official Comment 2 for 3VWs-Q1 — Independent value and efficiency of latent-query fusion

Related critical weakness. “The latent-query interface’s independent efficiency advantage remains unclear. Table 2’s no-stacking comparison gains 1.03 mIoU points at 4.58 times the payload. A comparable-budget attribute baseline with SIC and stacking would better isolate its contribution.”

Question. “What gains do latent queries provide over explicit attributes when both use the same SIC selection and unmatched-Gaussian stacking mechanisms under comparable communication budgets?”

Response: We corrected communication-volume (CV) accounting for compressed GSCollab variants and reran validation. The updated main results appear in our separate official comment, “Corrected CV and updated main results.”

The submitted no-stacking comparison leaves the efficiency question open. We trained the requested decoded-attribute + SIC-T + unmatched-Gaussian stacking control. All three rows below use one collaboration round, SIC-T during training and inference, and all 1,980 Semantic-OPV2V validation frames. Each model was trained separately. CV includes the selected token payload and packed ego-map request.

Transmitted / fused contentStackBinary IoU (%)mIoU (%)CV (MiB)
Decoded attributes / attributes, 96-D backbone✓73.2941.190.206
Pose + 96-D query / Latent Queries✓72.8742.640.695
Pose + 17-D query / Latent Queries✓72.3842.260.189

Latent fusion uses attention over spatially matched queries to update the ego Gaussians; the attribute control decodes Gaussians before explicit attribute fusion. With a 96-D backbone, SIC, and stacking in both settings, latent fusion gains 1.45 mIoU points but uses more CV and has 0.42 lower binary IoU. That comparison is not at a matched budget. A 17-D transmitted query gives a 24-float token, the same length as a decoded-attribute token (7 pose values and 17 attributes). At similar CV, it reaches 42.26 versus 41.19 mIoU (0.189 versus 0.206 MiB), with lower binary IoU (72.38 versus 73.29). Separately trained models can select different numbers of tokens, so equal token length need not give equal total CV. The 17-D result favors the complete compact-latent design on mIoU at this operating point; the backbone and fusion operation also differ, so the transmitted representation is not isolated.

Additional experiments are underway. We will present their accuracy and measured CV in one ablation table in a single shared follow-up official comment.

AuthorsOpenReview ↗

Official Comment 3 for 3VWs-Q2 — Handling pose information and cross-frame consistency

Related critical weakness. “Cross-frame query consistency lacks direct validation. Separate pose and attribute encoders do not guarantee pose-independent queries, and existing single-agent ablations and pose-noise tests do not directly verify consistent recovery of frame-invariant attributes after coordinate transformation.”

Question. “How is pose information handled in the final latent queries? Does re-expressing the same token in another coordinate frame preserve its decoded frame-invariant attributes and semantic predictions?”

Response: Decoded attributes are not strictly invariant: the receiver reconstructs them using the transformed pose, and training does not enforce exact preservation.

1. How is pose handled in the final latent queries? At the sender, each token carries a compressed partial query together with its Gaussian center and orientation in sender coordinates. The query includes embeddings of scale, opacity, and semantic scores; pose remains explicit so that each receiver can transform the token into its own frame without requiring a separately pose-conditioned query for every receiver.

At the ego, each received center and orientation is transformed into ego coordinates. The transformed pose is encoded and added to the decompressed partially attribute-embedded query, completing the attribute-conditioned query. For matched tokens, neighbor and ego queries enter attention. The fused ego query incorporates collaborator evidence from the full set of Gaussian attributes and is decoded with the ego pose embedding by the Gaussian-update decoder shared with single-agent refinement. Collaboration thus follows the same latent-to-Gaussian update pathway.

For unmatched tokens, a dedicated decoder reconstructs scale, opacity, and semantic scores using the fully attribute-embedded query; the transformed center and orientation are retained when these tokens are stacked with the updated ego Gaussians for occupancy prediction.

2. Does changing frames preserve the decoded attributes? We used a frozen GSCollab-112T checkpoint for this inference-only audit. Across all 1,980 Semantic-OPV2V validation frames, we recorded 5,749 neighbor-to-ego rigid transforms, one per neighbor in each frame. After filtering for centers inside the receiver's range and original opacity of at least 0.01, 2,938 neighbor pairs in 1,679 frames contributed about 17.68 million eligible primitives per run. For each eligible primitive, we compared its original single-agent Gaussian scale, opacity, and semantic scores before compression with attributes reconstructed from its decompressed query by the receiver's decode-and-stack attribute decoder, conditioned on the transformed ego-frame pose.

Attribute comparisonResult
Scale, relative L1 / absolute MAE0.17093 / 7.00 cm
Opacity MAE0.27443
Semantic-label agreement84.14%

Scale errors average the relative or absolute difference across three axes; opacity MAE uses absolute difference, and semantic agreement compares the highest-scoring labels. We first average eligible primitives within each frame, then average frames equally. A same-GPU repeat gave stable results across all three metrics.

The recorded transforms averaged 37.04 m in translation and 84.81° in absolute yaw. Most eligible primitives retain the same semantic label, while scale and opacity show nontrivial differences. This comparison includes compression and reconstruction effects as well as the frame change.

The dedicated attribute decoder reconstructs unmatched tokens for stacking; a separate Gaussian-update head handles matched-token fusion. Reconstruction MSE and occupancy supervision encourage useful outputs from the attribute decoder, without enforcing exact recovery of the original Gaussians. Here, "frame-invariant attributes" means that scale, opacity, and semantics require no explicit rigid-coordinate transform. Their decoded values can still change with the pose used for reconstruction.

3. Does this variation harm collaborative occupancy? We compared two inference paths using the same frozen GSCollab-112T checkpoint on the same Semantic-OPV2V validation split. The standard path reconstructs the attributes of selected unmatched stacked tokens from the decompressed query at the transformed ego-frame pose. The ideal-preservation control instead carries their original sender Gaussian scale, opacity, and semantic scores into the ego frame; centers and orientations are transformed normally, and matched-token fusion is unchanged.

Unmatched-token attributesBinary IoU (%)mIoU (%)
Receiver-frame reconstruction (standard)72.48341.818
Original sender attributes (ideal preservation)71.82741.881

In this fixed-seed full-split comparison, the standard receiver-frame reconstruction improves binary IoU by 0.656 points over retaining the original sender attributes, while mIoU is 0.063 points lower.

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

This paper introduces Gaussian splats as a medium for cooperative occupancy prediction. The proposed approach encodes multi-view images into Gaussian representations, selectively share those primitives, and use 3D transformations to establish vicinity before fusion within a neighborhood. Evaluations on two datasets show that proposed framework is superior than two existing baselines.

Critical strengths
  • GS as medium makes a lot of sense, the design of a cooperative framework around is a solid effort.
  • Evaluation shows consistent improvement over existing methods, especially lauded effort of evaluation under different bandwidth constraints.
Critical weaknesses
  • Evlauation is a little thin beyond main results, ablation, attribution is missing
  • Communication overhead and compression is under-specified.
Questions for authors

The evaluation is a little thin.Especially on sensitivity to pose errors and communication latencies. Because GS are spatially anchored, it would be critical to evaluate the sensitivity to transformation errors between transmitter and receivers, either due to pose error or delay.

Ablation wise, the key missing control is decoded attributes + stacking + SIC-T. Please also clarify whether the reproduced VOGS-L merges unmatched neighbor Gaussians and how this compares to the original VOGS-CP fusion. Without this control, it is unclear how much of the improvement over VOGS-L comes from transmitting latents at all.

Communication overhead needs to be explicitly accounted for. As a community making cooperative perception towards a mature technology, we should pursue cooperative gains under realistic wireless capacity.

Also the data request and delivery mechanism introduces additional delays, which is not evaluated. It’s debatable whether that’s a good mechanism in practice.

7 comments
AuthorsOpenReview ↗

Official Comment 1: Corrected CV and updated main results

We first provide corrected communication-volume (CV) results for PEQC-compressed GSCollab in the manuscript's table layout for easy comparison. We address the reviewers' questions in subsequent official comments. PEQC CV previously counted 128-dimensional queries reconstructed after reception rather than the compressed queries transmitted. The code correction changes byte accounting, not the trained models or prediction logic. We reevaluated the reported settings on the full validation split using their existing checkpoints; pose and packed SIC-request bytes remain included. Independent evaluations can yield different IoU/mIoU, so each updated accuracy value is reported with the CV measured in the same evaluation. Accuracy changes relative to the submission come from reevaluation, not from the byte-count correction. Unchanged baselines retain their manuscript values. The central result remains: compressed SIC settings exceed VOGS-L in semantic mIoU at lower CV.

CV is MiB per collaborative frame; IoU/mIoU are percentages. Slash-separated cells use CV/IoU/mIoU. N: no SIC; I: SIC at inference only; T: SIC in training and inference. ✓: stacking; ✓→--: stacking only during training. GSCollab-128 omits PEQC.

Table 2. Semantic-OPV2V fusion and communication controls

SettingFused contentStackProtocolIoUmIoUCV
VOGS-LDecoded attributes--N71.5438.900.941
GSCollab-128Latent queries--N72.8839.715.113
GSCollab-128Latent queries✓N74.1142.835.186
GSCollab-112Latent queries✓N73.9243.194.828
GSCollab-112Latent queries✓I73.2242.470.914
GSCollab-112Latent queries✓T72.4742.030.773
GSCollab-96 backboneDecoded attributes--T71.8638.910.196
GSCollab-96Latent queries✓→--T72.2939.900.695

Table 5. Co3SOP operating points (CV/IoU/mIoU)

MethodStackNIT
VOGS-L--0.786/77.74/41.19----
GSCollab-128--4.903/77.31/42.56----
GSCollab-128✓4.460/76.18/43.930.929/76.53/43.100.909/77.17/43.44
GSCollab-64✓2.533/78.06/44.440.523/77.94/43.900.511/77.89/44.45
GSCollab-80✓3.164/77.81/45.840.508/77.65/45.350.504/78.04/44.86
GSCollab-96✓3.884/77.63/44.970.613/77.51/44.540.566/77.71/44.67
GSCollab-112✓4.421/76.87/46.050.723/77.33/45.390.668/78.01/44.05

Table 8. SIC versus entropy-confidence gating

WidthHeur. mIoUSIC mIoUGainHeur. CV (MiB)SIC CV (MiB)CV reduction
12840.31841.501+1.1831.5761.10529.9%
6440.71241.938+1.2260.7630.51732.3%
8040.26340.913+0.6501.0600.73330.9%
9640.22041.618+1.3981.0700.89016.9%
11240.97142.474+1.5031.3440.91432.0%

Table 9. Semantic-OPV2V operating points (CV/IoU/mIoU)

MethodStackNIT
VOGS-L--0.941/71.54/38.90----
GSCollab-128--5.113/72.88/39.71----
GSCollab-128✓5.186/74.11/42.831.105/73.15/41.500.942/73.47/41.92
GSCollab-64✓2.923/74.43/42.490.517/73.67/41.940.484/73.65/41.47
GSCollab-80✓3.397/73.37/41.930.733/72.51/40.910.591/72.44/41.68
GSCollab-96✓4.022/74.69/42.340.890/73.73/41.620.695/72.87/42.64
GSCollab-112✓4.828/73.92/43.190.914/73.22/42.470.773/72.47/42.03

Table 10. Semantic-OPV2V erasure sweep (CV/IoU/mIoU)

Erasure rateVOGS-LGSCollab-64TGSCollab-80TGSCollab-96TGSCollab-112T
0%0.941/71.54/38.900.484/73.65/41.470.591/72.44/41.680.695/72.87/42.640.773/72.47/42.03
5%0.894/71.56/38.780.463/73.64/41.350.564/72.45/41.520.662/72.84/42.470.736/72.45/41.89
10%0.847/71.53/38.710.440/73.59/41.220.535/72.41/41.690.629/72.82/42.360.699/72.44/41.86
20%0.753/71.54/38.710.396/73.52/40.790.480/72.36/41.150.563/72.76/42.100.625/72.38/41.51
30%0.659/71.51/38.490.351/73.40/40.610.425/72.30/40.980.498/72.65/41.750.552/72.28/41.07
50%0.471/71.40/38.150.262/73.08/39.860.315/72.02/40.560.366/72.36/40.890.405/72.03/40.65

Table 10, CoHFF feature-sparsification reference (unchanged)

Metric0%50%80%95%99%
CV (MiB)15.7647.8873.1570.7920.162
IoU50.4649.5649.5348.5248.02
mIoU34.1632.9732.7030.1329.48
AuthorsOpenReview ↗

Official Comment 2 for bVkA-Q1 — Pose errors and communication latency

Related critical weakness. “Elvauation is a little thin beyond main results, ablation, attribution is missing”

Question. “The evaluation is a little thin.Especially on sensitivity to pose errors and communication latencies. Because GS are spatially anchored, it would be critical to evaluate the sensitivity to transformation errors between transmitter and receivers, either due to pose error or delay.”

Response: We corrected communication-volume (CV) accounting for compressed GSCollab variants and reran validation. The updated main results appear in our separate official comment, “Corrected CV and updated main results.”

Appendix Figure 7 compares GSCollab and VOGS-L under zero-mean Gaussian position noise of 0.1–0.8 m and heading noise of 0.1–0.8°. To test delayed observations separately, we evaluated fixed temporal staleness on Semantic-OPV2V: the ego observation and ground truth remain at time t, while each neighbor observation comes from t-d. Alignment uses the neighbor pose at t-d and ego pose at t. At the dataset's 10 Hz sampling rate, one frame is about 100 ms. Both methods use fixed weights and the same 1,935 validation frames with five preceding frames.

MethodDelay (frames)mIoU (%)Vehicle IoU (3D, %)Vehicle IoU (BEV, %)CV (MiB/frame)
VOGS-L038.6660.8475.950.952
VOGS-L138.5660.0375.250.950
VOGS-L338.5358.5173.350.948
VOGS-L538.4957.7272.180.946
GSCollab-112T041.5264.3677.690.780
GSCollab-112T141.5363.4276.850.783
GSCollab-112T341.4161.8975.030.790
GSCollab-112T541.3961.2373.890.794

At five frames (about 0.5 s), GSCollab loses 0.13 mIoU, 3.13 points of 3D Vehicle IoU, and 3.80 points of BEV Vehicle IoU relative to zero delay; VOGS-L loses 0.17, 3.12, and 3.77 points, respectively. GSCollab has higher absolute mIoU and Vehicle IoU at every tested delay. The zero-delay values here use a common 1,935-frame subset; the main results use all 1,980 validation frames. BEV Vehicle IoU is the projected 2D metric from Appendix Table 6, distinct from 3D Vehicle IoU. This test uses correct historical poses; it does not model packet-arrival jitter or time the request–response exchange. Q3–Q4 address the latter's timing.

AuthorsOpenReview ↗

Official Comment 3 for bVkA-Q2 — Missing decoded attributes + stacking + SIC-T control

Related critical weakness. The reviewer's concern about limited ablation and missing attribution is quoted under Q1 above.

Question. “Ablation wise, the key missing control is decoded attributes + stacking + SIC-T. Please also clarify whether the reproduced VOGS-L merges unmatched neighbor Gaussians and how this compares to the original VOGS-CP fusion. Without this control, it is unclear how much of the improvement over VOGS-L comes from transmitting latents at all.”

Response: We trained the requested decoded attributes + unmatched-Gaussian stacking + SIC-T control and evaluated it on all 1,980 Semantic-OPV2V validation frames. All rows below are separately trained, unlike the manuscript's inference-only stacking ablation. The first four rows use a GSCollab-96 backbone; the last uses a 17-D backbone. All use 12,800 initial Gaussians, a 0.4 m matching radius, SIC-T during training and inference, and one collaboration round. IoU/mIoU are percentages; CV includes the selected Gaussian payload and packed SIC request map.

Transmitted and Fused ContentStackIoUmIoUCV (MiB)
Shared Refinement Head Decoded Attributes; Attribute Fusion--71.8638.910.196
Shared Refinement Head Decoded Attributes; Attribute Fusion✓73.2941.190.206
Pose + 96-D Latent Query; Latent Fusion--72.5140.860.754
Pose + 96-D Latent Query; Latent Fusion✓72.8742.640.695
Pose + 17-D Latent Query; Latent Fusion✓72.3842.260.189

The previously reported 72.29 IoU / 39.90 mIoU / 0.695 MiB result switches stacking off only at inference on a stacking-trained model. The newly trained stacked decoded-attribute model reaches 41.19 mIoU, versus 38.91 without stacking. With stacking enabled, 96-D latent fusion reaches 42.64 versus 41.19 mIoU for attribute fusion, but uses more CV (0.695 versus 0.206 MiB) and has lower binary IoU (72.87 versus 73.29). The 17-D latent token matches the decoded-attribute token's 24-float length and reaches 42.26 mIoU at 0.189 MiB, versus 41.19 at 0.206 MiB for the stacked attribute control; binary IoU is lower (72.38 versus 73.29). This near-budget result favors the complete compact-latent design on mIoU. The 17-D model also changes the backbone and fusion operation: it applies cross-attention to latent queries, while the attribute control aggregates decoded Gaussian updates.

Our reproduced VOGS-L row in Table 2 sends decoded Gaussian attributes, uses neither SIC nor stacking, and applies the neighborhood-based attribute refiner to matched ego–neighbor Gaussians. The new decoded-attribute control uses the same matched-Gaussian attribute refiner, but SIC-T selects transmitted Gaussians and selected unmatched Gaussians are stacked; VOGS-L excludes unmatched received Gaussians from its final output. The VOGS-CP paper separately reports zero-shot and naive variants that concatenate received Gaussians.

Additional experiments are underway. We will present their accuracy and measured CV in one ablation table in a single shared follow-up official comment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.