acceptodds
Withdrawn from ICLR 2027

Neural Implicit Probability Fields for Lossless Pointmap Compression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Pointmaps have recently emerged as a new geometric output of feed-forward 3D reconstruction models, creating a new target for 3D data compression. Their high spatial resolution and coordinate precision, however, incur substantial lossless storage costs. We propose Neural Implicit Probability Fields, a lossless pointmap codec that uses an implicit neural representation (INR) as an entropy model rather than a signal reconstructor. From a compression perspective, backpropagation can be viewed as compressing pointmap-specific statistical structure into the parameters of a compact neural probability field. To make this idea practical, we combine reversible geometric prediction with neural probability modeling, allowing the INR to focus on the remaining uncertainty rather than reconstructing the signal itself. The resulting codec preserves the native pixel-to-point correspondence and supports exact 16-bit reconstruction, while explicitly accounting for the transmitted neural model. Across diverse domains relevant to world modeling and embodied intelligence, our method reduces BPP by 1.4–30.3% over the strongest grid-preserving baselines, demonstrating that neural optimization can provide an effective mechanism for lossless compression of high-precision 3D geometry.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Pointmap, Implicit Neural Representations, Neural Compression

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_A95SOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors proposed to encode the pointmap of a 3D geometry via implicit neural representation. A neural network is trained to predict the the pointmap values while the remaining residual is entropy encoded. The implicit neural network training is done with different optimizers and with 8-bit quantization of the network itself. The authors reported modest compression of the saved information across multiple different datasets, with the rate of the compression varies between them.

Critical strengths

The paper attempted to perform pointmap compression with a neural network, while implicit neural network compression isn’t a new field by itself, the compression of a pointmap is a relatively new task, if done well, can be beneficial for 3D geometry representation.

The selected evaluation set is relatively thorough with different scenes included.

Critical weaknesses

The empirical performance gain (compression rate) is very limited and only works in certain datasets, the rest has nearly no improvement.

The authors did discuss the encoding and decoding time related to the compression and decompression. However, the reported numbers does seem to be reasonable given today’s library and hardware. See more in the questions.

Questions for authors

Do the authors have further analysis on why some datasets are more compressible than the others. More specifically, why the proposed method works better on some datasets than the others.

The proposed method also includes the entropy encoding/decoding part, does the measured time include that as part of the process.

For the baseline comparison, JPEG XL as example, what kind of library, setting, thread, hardware (with dedicated hardware media codec used or not) leads to the reported values. The author reported encoding and decoding speed for the images are a bit too slow, particularly considering how low resolution the images are.

Reviewer_UD2xOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses lossless compression problem of pointmaps, a newly mainstream geometric output of feed-forward reconstruction models. Specifically, the proposed Neural implicit probability fields based lossless pointmap compression framework combines reversible geometric prediction with neural probability modeling, allowing the INR to focus on the remaining uncertainty rather than reconstructing the signal itself. Experimental results demonstrate that the efficiency of the proposed method over the strongest grid-preserving baselines on lossless compression task of high-precision 3D geometry.

Critical strengths
  • Innovative research task. Pointmaps are newly mainstream, and I am not aware of a published dedicated lossless pointmap codec. The motivation (image codecs ignore irregular masks; point-cloud codecs lose pixel alignment) is accurate and well-argued.
  • Conceptual contribution. Reframing the INR from signal reconstructor to entropy model / probability field, with the compression-oriented reading of backpropagation, gives the paper a coherent and memorable narrative anchor.
Critical weaknesses
  • Incomplete overview of related work. Related work coverage omits the depth map compression line, which undermines the novelty claim. Pointmaps are obtained by back-projecting depth with calibrated intrinsics (Appendix A.3, Eq. 22) , i.e., pointmaps and depth maps are the same underlying geometry. Yet Related Work contains no depth-map compression reference at all. However, there is a substantial, mature literature on high-precision depth-map lossless compression (e.g., end-to-end lossless compression of high-precision depth maps guided by pseudo-residual; lidar range-image compression via deep delta encoding), with problems almost structurally identical to the 16-bit high-precision setting here. Since pointmap = depth + intrinsics, what are the differences and advantages compared to depth graph compression?

  • Experimental setting on comparison with grid-preserving baselines is unfair. As stated in Appendix A.4, PNG/JPEG2000/JPEG XL encode the complete XYZ raster including values at invalid locations, whereas the proposed method codes only valid pixels (mask counted separately). This setting systematically inflates baseline bitrates in the paper's favor. On Robot (2.4%) and Internet (1.4%), the reported advantage is small to begin with; without a mask-aware baseline (coding only valid pixels), it is unclear whether the advantage survives.

  • Incomplete INR method baseline comparison. The closest INR-based baselines are absent from the experiments, and the novelty axes are not sharply delineated. Related Work cites LINR-PCGC, HybridINR-PCGC, LosslessINR, FNLIC, Cool-chic — but no INR-based method appears in Table 2 (only traditional image codecs and G-PCC/NBP). Given the paper's claim is precisely "INR as a probability field, not a reconstructor," a quantitative or semi-quantitative comparison/discussion with the nearest INR-line works (e.g., LosslessINR on the image side, LINR-PCGC on the point-cloud side) is needed, or an explicit statement of why they are not comparable. Additionally, the novelty axes vs. LosslessINR (INR reconstructs bit-plane values) and FNLIC/Cool-chic (per-instance fit + entropy model, but on natural images and mostly lossy) should be decomposed item by item (object of modeling / modeling target / mask handling / lossless-vs-lossy), not dispatched in one sentence at the end of §2.1.

  • Encoding latency sits in tension with the stated application motivation. At 26.61–28.45 s/frame (800 steps), per-frame encoding cost is high for the world modeling and embodied intelligence scenarios that motivate the paper (continuous accumulation, low latency). The paper acknowledges the trade-off but, given the motivation is a load-bearing part of the framing, should offer a more concrete mitigation direction or stage the conclusion more carefully.

Questions for authors
  • If the grid-preserving baselines were made mask-aware (coding only valid pixels), how much of the Robot/Internet advantage would remain?
  • Why is no INR-based method included in the experimental comparison, given the Related Work emphasizes the INR lineage?
  • How does the method behave as pointmap resolution grows (e.g., toward the 2K–4K regime used by modern world models), where per-pointmap INR overhead may dominate?
Reviewer_kQH3OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper combines a mask-aware MED transform, bit-plane arithmetic coding, and a per-pointmap implicit neural probability model for 16-bit pointmap compression. The reversible transform and decoder synchronization are described relatively clearly. The evaluation covers four domains and includes useful component ablations. A mask-aware MED predictor produces reversible residuals, which are encoded from the most to the least significant bit plane using a four-stage spatial order. A compact per-pointmap SIREN predicts bit probabilities from coordinates and decoded context. Its fitted parameters are quantized, coded, and transmitted, and the reconstructed model is used on both sides. The reversible coding framework is interesting, but the paper needs a clearer definition of end-to-end losslessness, more rigorous bitrate accounting, stronger direct baselines, and statistical analysis.

Critical strengths

A complete coding formulation with explicit adaptation costs. The integer transform is bijective, the context order is decodable, and transmitted model and mask bits are counted.

Useful compression gains and transparent cost analyses. At 800 fitting steps, the method has lower mean bitrate than the tested grid-preserving alternatives in all four domains, including 5.56 to 3.88 bits per valid pixel on Human

Critical weaknesses

The main concerns are that the claimed “losslessness” appears to apply mainly to the already quantized 16-bit pointmap, while the quantization error and the bitrate of the quantization parameters are not sufficiently specified. The BPP normalization uses the number of valid points as the denominator, while the baselines still encode invalid pixels, which may affect fairness.

The paper also lacks strong direct baselines such as FNLIC, LosslessINR, MED combined with ArIB-BPS, ordinary MLP/CNN probability models, and mask-aware codecs. The default encoding time is approximately 26–28 seconds per frame, which substantially limits practical applicability, and the small improvements of 1.4%–2.4% are reported without sufficient variance or confidence intervals.

The distinctive coordinate-field contribution is not isolated. A comparable context model without coordinate inputs and a conventional adaptive entropy model using the same transform would reveal which gains require the proposed field.

Questions for authors

How does a matched context model without coordinate inputs perform, and how does the same reversible transform perform with a conventional adaptive entropy coder?

For depth-derived pointmaps, what is the complete rate for depth/Z, necessary metadata, and exact X/Y corrections, verified against the identical quantized XYZ target?

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

Related papers

Discussion (0)

Sign in to comment.