acceptodds
Withdrawn from ICLR 2027

DECISION-FLIP: BUDGET FREE STRUCTURED PRUNING FOR GAUSSIAN OCCUPANCY MODELS

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Occupancy predictors rely on large persistent banks of Gaussians or learned queries whose size is fixed by hand. Common pruning proxies, such as opacity, activation frequency, or parameter magnitude, do not measure whether removing a row changes the model's voxel decisions. We introduce Decision-FLIP, a budget-free compilation framework whose scoring and structure selection need only unlabeled calibration inputs. Exact Mean-FLIP scores each persistent row by the number of voxel argmax decisions that change when the row is removed from the readout; a guarded Otsu split of the log-score distribution selects the survivor count without a target and abstains when the split is unsupported; and the survivors are physically compiled into a smaller checkpoint, so camera sampling, attention, and aggregation shrink with the bank. For streaming banks, scores are integrated over each row's lifecycle through warp, death, and refill. Across four Gaussian-bank model-dataset settings and one persistent-query transfer, the automatically selected cuts remove 12–79% of eligible rows and cut bank/query-dependent latency by 7–45% across settings with component profiles, with whole-model savings of up to 8.0% on the primary endpoints bounded by fixed-cost image, voxel, and sensor modules and peak-memory savings of up to 16.6%. Under matched one-epoch recovery controls, semantic mIoU changes by at most 0.18 point. Same- controls show that the ranking, not only the bank size, matters: random removal collapses GaussianFormer-1 and OccFormer, while label-free Mean-FLIP matches or outperforms supervised Taylor ranking on all five settings where that control was run. The method applies when removable primitives keep a stable checkpoint-row identity; it is budget-free in survivor count, not parameter-free.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Structured Pruning, Model Compression, 3D Semantic Occupancy Prediction, Gaussian Scene Representation, Autonomous Driving, Automatic Sparsity Selection, Inference Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_ahS9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Decision-FLIP, a structured pruning method for occupancy predictors that keep a persistent bank of Gaussians or learned queries. Each bank row is scored by how many voxel argmax decisions change when its contribution is removed from the (additive) readout, using unlabeled calibration frames. For streaming models like GaussianWorld, the score follows each row through warp, death and refill. The number of survivors is chosen by Otsu thresholding on log scores, with a separation gate that keeps the full bank if the split looks weak. The selected rows are then physically removed from all aligned tensors, producing a smaller checkpoint that loads strictly.

Experiments cover GaussianWorld, GaussianFormer-1 and GaussianFormer3D on SurroundOcc, a self-trained GaussianFlowOcc on Occ3D, and OccFormer on SemanticKITTI as a query-bank transfer. GaussTR serves as a negative control. Reported results:

  • 12–79% of eligible rows removed;
  • 7–45% lower bank-dependent latency and 2.4–8.0% lower whole-model latency;
  • semantic changes within about 0.18 point after one matched recovery epoch.
Critical strengths

1. The criterion is simple, exact and label-free, and the controls are careful. Counting actual argmax flips under an exact readout deletion ties importance directly to the output that matters. The paper supports it with matched full-vs-compiled recovery, same-K ranking controls, strict-load compilation, and an informative negative control: on GaussTR the gate passes, but the ranking is no better than random, because rows have no stable identity. I appreciated that this failure is reported rather than left out.

2. The reduction is physical and the reporting is transparent. Rows are actually removed from every aligned tensor, so latency and memory numbers come from smaller real checkpoints rather than masks. Component-level timing, peak memory and per-class results are all there, and the limitations (Amdahl bounds, the row-identity requirement) are stated openly.

Critical weaknesses

(1) The practical gain is small, and it's hard to attribute it to the proposed score. End to end, the primary endpoints save 2.4–8.0% latency. Peak memory goes down on some models but is flat or slightly up on others (GaussianFormer-1 +0.8%). Amdahl's law explains this, but it leaves a fairly thin deployment case.

More importantly, outside OccFormer, the flip score barely beats the cheap proxies a practitioner would try first:

  • GaussianFormer-1, before recovery: opacity loses 0.183 mIoU vs. 0.157 for the flip score. After recovery at K=17,296K=17{,}296, the two are essentially identical (19.105 vs. 19.108).
  • GaussianWorld: opacity −0.244 vs. −0.179.
  • GaussianFormer3D: magnitude −0.071 vs. −0.044.

Gaps of a few hundredths of a point don't convince me that the decision-level score is worth its extra complexity over opacity. The budget selection (Otsu with a gate) is a standard tool, and the paper itself presents it as Pareto-competitive rather than better than GMM or Kneedle. So I'm not sure what a practitioner gains over a simple opacity cut.

(2) The scope is narrow, and the headline accuracy claim depends on a recovery step that can hide differences. The method needs persistent rows with stable identity and an exactly decomposable additive readout. That rules out models like GaussTR and essentially limits it to a small family of Gaussian/query-bank occupancy models, which makes the broader impact for ICLR feel limited.

The "within 0.18 point after matched recovery" claim also mixes quite different situations:

  • On OccFormer, the recovery epoch makes both arms worse (full 13.50 → 12.99, compiled 13.34 → 13.02), so "+0.03 after recovery" compares two models that are both below the released checkpoint.
  • On GaussianFormer-1 and GaussianFormer3D, the same epoch moves absolute mIoU by +1.7 and +2.7 through pipeline adaptation, which is far larger than the compilation cost being measured.

In these cases, the post-recovery deltas don't add much beyond the zero-recovery numbers.

Questions for authors

1. Would Otsu on opacity scores do about as well? Since opacity is close to the flip score on GaussianWorld and GaussianFormer-1, could you apply the same guarded Otsu rule to opacity (or magnitude) scores and report the selected K and accuracy before and after recovery? That would show whether the benefit comes from the decision-level score or from the budget-free selection rule.

2. Which checkpoints are you starting from? GaussianWorld is compiled from an "opacity-calibrated" source checkpoint (21.902 mIoU) rather than the released one (21.759). What exactly is calibrated, and does the same step help the full model on its own? For OccFormer, given that the recovery epoch hurts both arms, would you consider making the zero-recovery comparison the primary result?

3. Are the bank-dependent latency numbers measured the same way across models? For GaussianWorld and GaussianFormer-1, Table 1 seems to report the Gaussian-side subtotal (encoder plus head). For GaussianFormer3D, the −13.2% appears to cover the encoder only. Including the head would give roughly −9.7%. Could you make the definition consistent across rows?

Reviewer_671VOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper is not a new occupancy model; it is a way to slim down an already-trained one. The model has a fixed set of Gaussian or query “parts” that vote on the class of each 3D voxel. The paper removes one part at a time, recomputes the voxel predictions, and counts how many classifications change; parts that change little are treated as redundant. It then sorts the importance scores, automatically picks a cutoff (and skips pruning if the scores are too messy), actually deletes the selected parts from the checkpoint, and compares the full and slimmed models after the same one-epoch fine-tuning. It reports removing 12–79% of eligible parts, up to 45% faster part-dependent computation and 8% faster whole-model inference, with less than 0.18 point accuracy difference after matched recovery—but it only works when each part has a stable identity and the readout can be exactly recomputed after deletion.

Critical strengths
  1. The paper targets a real deployment cost in occupancy models with large persistent Gaussian or query banks, and it provides a concrete way to physically remove redundant rows from an already-trained checkpoint rather than only masking them at inference time.

  2. The importance score is tied to final voxel-level class decisions instead of proxies such as magnitude, opacity, or activity, and it does not require labels, which makes the criterion well aligned with what pruning needs to preserve.

  3. The empirical comparison is more careful than is common in this area: full and pruned models receive matched recovery budgets, same-K ranking controls are included, per-class preservation is checked, and a negative control is provided for architectures without stable row identity.

Critical weaknesses
  1. The paper makes a fairly simple procedure sound much more complicated than it is. In plain terms, it removes one part at a time, counts how many voxel classifications change, deletes the low-scoring parts, and fine-tunes. The heavy terminology and fragmented presentation make the method harder to follow than its actual complexity warrants.

  2. The “budget-free” claim is overstated. It does not require a target sparsity, but it still fixes an offset and a separation threshold, and the threshold only decides whether to prune at all. Once it passes, the cutoff still determines how many parts are removed. Calling this “budget-free” invites readers to expect something more parameter-free than it is.

  3. The method applies to a narrow class of models. It requires each part to have a stable identity, the final prediction to be exactly recomputable after removing a part, and the relevant tensors to be aligned so they can be sliced together. Models that resample or reorder parts on every input are excluded. The paper uses a negative control to show this, but the abstract and introduction do not make the limitation salient enough.

  4. The writting is hard to follow. The paper introduces many terms and defensive qualifications, and the appendices are very long. A clearer statement of the procedure and its assumptions early in the paper would substantially improve readability and make the contribution easier to assess.

While the problem is practical and the evaluation has some careful elements, the technical contribution is limited and is presented with terminology that makes it appear more novel than it is. The “budget-free” claim is overstated, the applicability is narrow, and the empirical evidence is not strong enough to support the paper’s broader claims. I recommend rejection.

Questions for authors

Please see weakness

Reviewer_oY13OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Interesting paper but the presentation could be improved. The simple idea is that one can prune the learned Gaussian splat bank for occupancy prediction. The paper’s method works by figuring out a thresholding of which Gaussians to keep and which ones to discard, but instead of using the task loss, they score each bank row by how many voxel predictions flip when it is removed, on unlabelled frames, and then use Otsu thresholding on those scores to decide the cut.

It is a post-processing method where the threshold comes from the score distribution, and the pruned rows are sliced out to give a smaller checkpoint, followed by an optional one-epoch fine-tune. The learned bank itself is shared across scenes (only the generated Gaussians differ per scene), so a row’s importance is averaged over scenes. The pruning seems to be useful, losing about 0.2 mIoU from the full, unpruned model. However, the end-to-end speedup advantage is modest (2–8%).

Critical strengths
  1. Interesting application of Otsu’s thresholding method. This classical approach has been used in image processing for decades, but its use for pruning in semantic occupancy prediction is novel and interesting, and pairing it with a label-free measure of decision changes is a clean idea.
  2. Exhaustive experimentation on in-domain data. The method works across four Gaussian models plus OccFormer, and it handles the streaming case by tracking each Gaussian through warp/death/refill.
Critical weaknesses
  1. The practical value is limited. In my view, the bank is cheap to store in memory and not the bottleneck, so pruning it can only buy so much. The end-to-end latency gains are modest (2–8%) because the image backbone dominates, and the memory benefits are inconsistent across models. A smaller bank is useful only insofar as it buys runtime and memory, and here it buys relatively little.
  2. Pruning the learned bank in the general case (training on huge datasets) might hurt in long-tailed situations. The bank being pruned can be thought of as an amortized prior shared across every scene. My hypothesis is that as the training data grows and diversifies, more rows specialize to rare situations, and these are the rows with low average flip counts (and thus more likely to be pruned). It's difficult to believe that this method would generalize to long-tailed situations.
  3. Why not learn this pruning during training to get most of the advantage? This would also account for redundant rows that look unimportant individually but not jointly. Another way to think about it is to use compression directly to convert the bank slots into a smaller number of latents. Relatedly, how does the pruned model compare to a smaller bank trained from scratch?
Questions for authors

The method appears to be sound and well evaluated, but given the modest practical gains and open questions about long-tailed generalization, I lean toward rejection.

  1. Does the method hold up in long-tailed situations? Can the authors show results on rare subsets (e.g., night/rain scenes)? Evidence here would substantially strengthen the paper.
  2. Can the authors argue why this is not of limited value? The end-to-end gains are modest (2–8%). Is there a setting where pruning the bank matters more, e.g., optimized deployment, lighter backbones, or larger banks (maybe higher resolution), where the bank becomes a larger share of the compute?
Reviewer_UdM9OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper presents a pruning technique to reduce the number of learned Gaussian anchors in occupancy prediction networks. The authors propose a Mean-FLIP score that measures each anchor’s importance as the average number of voxel predictions that change when that anchor is removed, computed without labels. Using this score, they decide the anchors to keep or discard by applying otsu thresholding to the mean flip score. The discarded anchors are removed from the model, and the remaining anchors form a smaller model. Using decision-FLIP, the authors show that GaussianWorld, GaussianFormer, GaussianFormer3D, GaussianFlowOcc, Occformer can achieve latency and peak memory reductions of up to 8% and 16.6% respectively, with small loss in accuracy.

Critical strengths
  1. Idea is neat and simple, and computationally cheap.
  2. Authors provide a comprehensive evaluation.
Critical weaknesses
  1. Latency and memory reductions are modest.
  2. Does not compare with checkpoints trained on fewer gaussian banks
  3. The writing needs significant improvement.
Questions for authors

I like the simple idea presented in the paper. However, I believe the results are not strong enough. I have provided detailed comments below

  1. Results are not strong enough: The main objective of the paper, as I understand, is to prune Gaussian banks to improve on latency and peak memory required to run the model. However, the results in Table 1 show only small improvements. While the bank-dependent parts of the model get 23.3% faster (in Table A4), the end-to-end latency drops by only up to 8% compared to baseline, which is very modest, since the image backbone and other fixed-cost modules are not affected by pruning. The peak memory reduction is also in the single digits most of the time, and for GaussianFormer-1 it even slightly increases. So the question to the authors is this: Please provide a clarification of whether the main objective of Decision-FLIP is indeed to reduce latency and peak memory? Could the authors discuss settings where these gains would be more substantial?

  2. Would the benefits of Decision-FLIP (latency and memory reduction) be obtained by simply training the model with a smaller Gaussian bank of the same size? Is it possible to compare against such a baseline? I understand that this might require a lot of time along with a multi-GPU setup, and thus I’m not suggesting the authors run a training experiment. However, can the authors include a discussion on whether that is achievable instead? This is because models with smaller Gaussian banks are available publicly, which affects the motivation of the paper. For example, GaussianFormer-2 [1] has checkpoints at 6400, 12800 and 25600 Gaussians, and they measure the latency for these in Table 3 of their paper. The 12800 checkpoint reduces latency by about 10% with an mIoU drop of only 0.39, which is comparable to Decision-FLIP on GaussianFormer-1 (5.5% latency reduction, 0.17 mIoU drop). Can the authors apply decision-FLIP to the 25600-Gaussian GaussianFormer-2 checkpoint and compare it against the publicly available 12800 and 6400 checkpoints?

  3. The writing can be greatly improved. It was very dense and hard to read. Large portions of text do not contain necessary context required for a reader to understand. I had to really decipher several terms myself before I understood what the paper is trying to do. I have not given my score based on the writing quality for this review, but I strongly recommend the authors improve the paper's presentation. I have outlined a few points to fix here:

    a. The paper uses generator slots, bank rows, persistent rows, primitives, slots, anchors, queries, and Gaussians for roughly the same thing, interchangeably. Please use consistent terminology, and define terms when using them.

    b. Why use the term "compilation" in this context?

    c. The authors can significantly improve the introduction section by fixing the following issues:

    - Unintroduced terminology in the first paragraph such as "generator slots" (line 039) without defining what that is. The author should define things before using them in the text, or reference a section where it is defined.
    - When talking about sparse gaussian occupancy models (line 037), the authors need to cite representative sources to indicate what they are. I only got to know halfway through reading the introduction what the project is talking about.
     - Line 045 "Most structured pruning methods also require ..." needs a citation.
     - Line 047 "unlabeled calibration inputs" isn't defined, and what is "full class competition" in the same sentence? 
     - Line 083: What is "physical model compilation"? I don't believe this is a standard terminology, I do not see this term defined before as I read the introduction when I read this.
    

    d. It is recommended that the authors include a brief background section to introduce terminology and context to the reader. This would make several parts of the method section make sense. Here are a few things the author should fix:

             - Line 154: What is "stable mapping"? What do the authors mean by "stable" here?
     - "split-half stability" is undefined before usage (line 160).
     - Line 162: Include citations in place for GaussianWorld, GaussianFormer, etc. 
     - Line 215: The term non-recurrent bank is used without introducing it or a reference. I could later understand what that means from related work, but there should be some context for the reader to understand.
     - Section 3.4: A number of terms like "Death", "Lineage", "always alive banks" have to be clearly described before using them. 
    

[1] Huang, Yuanhui, et al. "Gaussianformer-2: Probabilistic gaussian superposition for efficient 3d occupancy prediction." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

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

Related papers

Discussion (0)

Sign in to comment.