acceptodds
Under review as a conference paper at ICLR 2027

BiOcc: Bidirectional Context Modeling for Semantic Occupancy Compression

Abstract

3D semantic occupancy has emerged as a unified representation of 3D scenes, with broad applications in scene generation, autonomous navigation, and collaborative perception. Nevertheless, transmitting dense 3D occupancy representations requires substantial bandwidth overhead, motivating efficient compression of both geometry and semantic labels. A natural approach is to treat semantic labels as point attributes and adapt existing point cloud codecs accordingly. Unlike conventional appearance attributes, semantic labels are spatially coherent and closely coupled with geometry, making them potentially informative for geometry coding. However, leveraging semantic information as context for geometry coding remains largely underexplored. To address this, we propose BiOcc, a bidirectional semantic occupancy compression framework that exploits early-decoded semantics for subsequent geometry coding and reconstructed geometry for finer-scale semantic coding. For semantics to geometry, label relative spatial context captures the arrangement of the same label neighbors, while the semantic category context encodes the current node's label. The semantic representations are combined with the causal geometry context to improve the prediction of occupancy patterns. For geometry to semantics, the local geometry context is combined with the available semantic context to predict label probabilities on occupied voxels. We further analyze the potential coding gain from semantic conditioning and show how semantic coarsening can reduce the geometry relevant information available for occupancy prediction. Extensive experiments on the outdoor nuScenes-Occ and indoor ScanNet-Occ datasets demonstrate that BiOcc achieves state-of-the-art performance in geometry and semantic compression.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Occupancy Compression, Entropy Modeling

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_KA2mOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents BiOcc, a bidirectional context modeling framework for semantic occupancy compression. BiOcc introduces a coarse-to-fine interleaved coding strategy that allows previously decoded semantic labels to assist subsequent geometry coding, while reconstructed geometry is further used to facilitate semantic coding. For semantic-conditioned geometry coding, the method combines label-relative spatial context and semantic category context with causal geometry features to refine the probability distribution over child occupancy patterns. Experiments on nuScenes-Occ and ScanNet-Occ demonstrate BiOcc achieves the state-of-the-art performance.

Critical strengths
  1. The core motivation is clear. Instead of treating semantic labels only as attributes to be compressed after geometry, this paper exploits them as informative context for geometry entropy modeling.

  2. The experimental results are convincing. BiOcc achieves strong BD-Rate performance on both outdoor and indoor datasets, and the ablation studies consistently support the effectiveness of the label-relative spatial context, semantic category context, and finer-level semantic conditioning.

Critical weaknesses

The comparison with several baselines requires substantial task-specific adaptation, including modifications to semantic representations, geometry reconstruction strategies, and semantic coding modules. A stronger controlled baseline using the same geometry backbone and coding pipeline as BiOcc, but without semantic-to-geometry conditioning, would help isolate the contribution of the proposed bidirectional context modeling.

Questions for authors

See weakness 1

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

This paper proposes BiOcc, an end-to-end AI-based codec that jointly compresses and reconstructs geometry and semantic labels. BiOcc performs compression and reconstruction based on a hierarchical octree structure, where geometry and semantic labels mutually benefit from each other's context to improve coding efficiency. Specifically, the decoded semantic labels at level L−1L-1 are used as context for geometry coding at level LL, while the reconstructed geometry at level LL is subsequently used as context for semantic label coding at the same level.

BiOcc also supports lossy reconstruction. Building upon the top-K selection-based reconstruction commonly used in AI-based lossy geometry compression, the proposed method selectively applies lossy reconstruction to a certain proportion of child occupancies, enabling adaptive control of geometry distortion and bitrate. Although the information-theoretic motivation behind BiOcc is interesting, the experimental methodology is not sufficiently explained, and the performance comparisons are not clearly presented. Consequently, it is difficult to determine whether the reported results reliably demonstrate the effectiveness of the proposed method.

Critical strengths

The context conditioning between semantic labels and geometry is mathematically formulated, and the causal relationships within the overall coding framework are clearly explained. In particular, BiOcc reports consistent coding gains across two different datasets.

Critical weaknesses

W1. Insufficient explanation of the performance metrics and experimental configurations in Table 1 and Fig. 4.

First, it is unclear how the rate points of BiOcc were generated. The paper does not specify at which octree levels top-K reconstruction is applied, how the values of K are transmitted, or whether the associated signaling costs are included in the reported bitrates. In particular, lossy compression allows various optimizations depending on the selection of rate points. However, no additional experimental results are provided to show how coding performance varies with λG\lambda_G. Therefore, it is difficult to determine whether the comparisons were conducted fairly.

Furthermore, the selection of rate points for the competing methods is also unclear. TopNet is described as using the top-K reconstruction mechanism adopted from BiOcc. However, as with BiOcc, the procedure for selecting its rate points is not explained, making it difficult to assess the fairness of the comparison. UPCC provides separate quality controls for occupancy and semantic labels, but the paper does not sufficiently explain how these controls were configured to generate the reported rate points. For G-PCC and AnyPCC, which appear to be primarily lossless coding methods, the paper does not explain how their lossy operating points were generated. Moreover, although most geometry codecs have been developed and optimized primarily using PSNR-based distortion metrics, neither D1 nor D2 PSNR is reported. This makes it difficult to verify whether comparable operating points were selected across the different methods.

Another major concern is that only BD-Rate values and RD curves are reported. The paper does not provide tables of the actual bpp, D1/D2 PSNR, IoU, or mIoU values. In Fig. 4, the RD curves of certain methods appear to be truncated, making it difficult to determine the operating ranges over which the BD-Rate values in Table 1 were calculated. In particular, the actual bpp values are essential for verifying the fairness of the comparisons against the bitrates reported in the original publications of the competing methods or under the MPEG Common Test Conditions.

W2. The sources of the reported coding gains are not adequately isolated.

Most of the compared methods support lossless geometry compression. In fact, this makes it relatively straightforward to conduct a fair comparison. The lossless bitrates can be directly compared at identical reconstruction quality. If RD curves are required, the same lossy preprocessing mechanism, such as identical voxel quantization or octree depth truncation followed by lossless coding, can be applied to all codecs.

Under such a configuration, all codecs would achieve identical IoU and mIoU values at each operating point, with only the bitrates differing. Therefore, the coding gains attributable to context conditioning should be evaluated separately from those attributable to the lossy reconstruction mechanism. Without this separation, it is difficult to clearly assess the effectiveness of BiOcc.

Questions for authors
  1. Could the authors describe the lossy mechanism and the parameters of every method in Table 1 and Fig. 4, and state whether each one is designed or tuned for IoU or for D1/D2 PSNR? In particular, for BiOcc, the octree levels where top-K is applied, how K is coded and counted, and the λ values. For G-PCC, the geometry lossy mode and its parameters. For AnyPCC, how the lossy points were obtained. For Unicorn, the configuration of the down-scaling and super-resolution path. For UPCC, how the occupancy and semantic quality settings were combined. Please also explain how BD-Rate is computed and give the integration interval for each row of Table 1.

  2. Please provide tables with the actual values at every operating point, including the lossless point, for every method and dataset. Each row should give the geometry bpp, semantic bpp, total bpp, IoU and mIoU. If the results of the compared methods in these tables differ from their published values, please explain why. In addition, please show plots over the full quality range drawn from these tables.

  3. Please report the lossless bitrate (geometry, semantic and total bpp) of all codecs. Since all compared methods would achieve identical reconstruction quality under lossless coding, these results would enable a direct comparison of their entropy modeling capabilities.

  4. For lossy compression, I suggest adding a comparison in which the same lossy preprocessing mechanism is applied to all codecs, so that the resulting RD curves differ only in bitrate at each reconstruction quality. If the authors retain different lossy reconstruction mechanisms for the compared methods, D1/D2 PSNR should also be reported to enable a more comprehensive evaluation of geometry reconstruction quality.

  5. I recommend including learned geometry codecs that work directly on LiDAR data and have public code, such as RENO (Real-Time Neural Compression for 3D LiDAR Point Clouds, CVPR 2025).

Reviewer_Pqu9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BiOcc for 3D semantic occupancy compression. The method interleaves geometry and semantic coding: decoded coarse semantics help predict finer geometry, and reconstructed geometry then helps predict semantic labels. The geometry probability model combines label-relative spatial patterns and category embeddings to refine occupancy predictions. The authors evaluate joint rate-distortion performance and conduct ablation studies on the nuScenes-Occ and ScanNet-Occ datasets.

Critical strengths
  1. The interleaved coding schedule is well motivated. Transmitting coarse semantics first makes them available for finer geometry prediction while preserving decoding causality. Reconstructed geometry subsequently provides context for semantic coding.

  2. Joint compression results on both datasets support the method’s practical value. Table 1 reports total-rate BD-Rate reductions of 70.32% and 87.61% relative to G-PCC on nuScenes-Occ and ScanNet-Occ, respectively, with better results than the evaluated adapted learning-based baselines.

Critical weaknesses
  1. The ablations do not clearly isolate the contribution of bidirectional conditioning. Tables 2 and 3 do not specify the rate and quality metrics or whether BiOcc-G retains the same coarse-semantic transmission schedule. A controlled total-rate comparison is needed to establish the benefit of bidirectional conditioning.
  2. Rate accounting and baseline adaptation limit the interpretation of the performance gains. The geometry comparison excludes the coarse-semantic bits required by BiOcc to reconstruct geometry. It therefore does not compare the complete recovery cost against geometry-only codecs. Several baselines also require task-specific adaptations, but the paper does not document their training, tuning, or operating-point selection well enough to assess whether they were adequately optimized. Besides, the definition of coarse semantics is inconsistent, and the connection between the theory and cross-level experiments needs clarification. Section 3.1 uses majority voting directly over finest-level voxels, whereas Appendix C.2 recursively votes over child-node labels.
Questions for authors
  1. Which rate and quality metrics are used in Tables 2 and 3? Does BiOcc-G retain the same semantic transmission schedule? Please provide a total-rate comparison between unidirectional and bidirectional variants with the same backbone and reconstruction strategy.
  2. Does the geometry compression advantage remain after including the required coarse-semantic overhead? Please provide the training settings, tuning ranges, and criteria for selecting operating points for the adapted baselines.
  3. Which coarse-semantic voting rule is implemented? Please reconcile the main text and appendix, and explain how the deterministic aggregation assumption applies to the actual contexts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.