H3DNAS: Hardware-Aware ONNX-Native 3D Point Cloud Model Compression
Anchit Mulye, Rhythm Baghel, Sujay Kumar Ingle, Hardik Jain
cs.LG · cs.AR · cs.NE
Quick verdict
H3DNAS prunes 3D point cloud models from ONNX alone, but its 'provable ceiling' claim is self-contradicted; CPU-only speedups and single-dataset, no-variance results cap conviction.
Summary
H3DNAS compresses pretrained 3D point cloud networks (PointNet, PointNet++, PointMLP) directly from ONNX files, without source code: a Channel Dependency Graph classifies operators into four constraint classes and computes a claimed compression ceiling ρf in linear time; a two-stage search combines L1 channel pruning, output-fidelity pre-screening on 32 random inputs, and optional GhostConv mutations under Jetson Orin Nano budgets, followed by onnx2torch-based fine-tuning. On ModelNet40 it reports 65.5%/43.2%/49.1% parameter reductions with −0.04/+0.08/−0.28 pp accuracy change and 1.99×/1.29×/1.67× CPU speedups, and the graph analysis flags hard-to-prune models (PCT, PTv3) before search. Main weaknesses: the ρf 'ceiling' claim is contradicted by achieved results (PointNet 65.5% > ρf 58.13%) and re-labeled a 'safety invariant' in the supplement; evaluation covers one dataset with no variance or significance testing; headline speedups are CPU-only — Jetson GPU gains (1.03×–1.45×, TRT FP16) are confined to the supplement.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- H3DNAS is the first source-code-free compression pipeline for 3D point cloud models, operating entirely on ONNX ModelProto graphs for search, mutation, and fine-tuning.
- The Channel Dependency Graph yields ρf, claimed in the abstract as a provable topological compression ceiling computable in O(|V|+|E|) — though the supplement re-labels it a safety invariant after PointNet's achieved 65.5% exceeds ρf=58.13%.
- On ModelNet40, PointNet/PointNet++/PointMLP are compressed by 65.5%/43.2%/49.1% parameters with −0.04/+0.08/−0.28 pp accuracy change and 1.99×/1.29×/1.67× CPU inference speedups.
- CDG analysis predicts before any search which architectures channel pruning cannot compress (PCT 12.34%, PTv3 13.97% ρf), redirecting to quantization or head pruning.
- Output fidelity on 32 random inputs is a robust label-free zero-shot proxy that remains informative where training-free proxies like SWAP-Score allegedly degrade.
Strengths
Pre-search feasibility analysis is genuinely useful and cheap
The CDG analysis (Sec. 3.2, Table 1) computes a free-parameter fraction from ONNX topology alone in O(|V|+|E|) in under a second, correctly flagging hard-to-prune architectures (PCT ρf=12.34%, PTv3 13.97%; MobileNetV2 with only 2/53 free nodes) before any search — a decision tool prior compression papers lack.
Clean problem framing with a label-free screening proxy
Output fidelity (logit cosine similarity on 32 random inputs) pre-screens candidates before labeled evaluation (Sec. 3.3), and the supplementary Table 15 shows a real screening gap: H3DNAS Stage 1 zero-shot 84.64% vs 32.54% for Random/L1 baselines at ~62% parameter reduction on PointNet.
Strong headline results with clear source-free framing
Honest GPU-valuation in the supplement
Broad stress-testing of the graph analysis beyond the three headline models
Weaknesses
The central theoretical claim is internally contradicted
Single dataset, no variance, gains within noise
Headline speedups are CPU-only while GPU gains collapse
Inconsistent measurements and a conclusion that does not match its table
Editorial and protocol inconsistencies undermine trust
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
H3DNAS — Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does H3DNAS do?
H3DNAS compresses 3D point cloud neural networks (PointNet, PointNet++, PointMLP) starting only from their ONNX model files — no original training source code is needed. It analyzes the ONNX graph with a 'Channel Dependency Graph' to find which layers can safely lose channels, searches for good pruned variants using L1 importance plus an output-similarity score on random inputs, optionally swaps convolutions for cheaper GhostConv blocks, and fine-tunes the result after rebuilding a trainable model with onnx2torch.
How much smaller and faster are the compressed models?
On ModelNet40 classification, H3DNAS reports parameter reductions of 65.5% (PointNet), 43.2% (PointNet++ SSG) and 49.1% (PointMLP) with accuracy changes of −0.04, +0.08 and −0.28 percentage points respectively. CPU (ONNX Runtime) inference speedups are 1.99×, 1.29× and 1.67×; on the Jetson Orin Nano GPU with TensorRT FP16 the measured speedups are much smaller: 1.06×, 1.03× and 1.45× (supplementary Table 14).
Is the search really label-free, and do the authors release code?
The search itself uses only the ONNX file plus unlabeled random inputs (output fidelity is cosine similarity of logits). However, the headline accuracy numbers come after 15 epochs of fine-tuning on labeled ModelNet40 via onnx2torch, which requires the dataset's labels but not the original model source code. The paper points to a public GitHub repository (ClarityLab-Org/h3dnas) and an anonymous.4open.science mirror with code, pre-compressed ONNX models, and a hardware-analysis script that reproduces latency/parameter/FLOPs/accuracy measurements.
What is the ρf 'compression ceiling' and does it hold?
ρf is the fraction of Conv/Gemm parameters in freely prunable nodes, computed from the ONNX graph topology alone in linear time. The abstract presents it as a provable compression ceiling, but the paper's own results exceed it (PointNet: 65.5% achieved vs ρf 58.13%), and the supplementary material re-labels ρf 'a safety invariant, not a compression ceiling' because pruning a free layer also removes downstream input channels, so realized reduction can surpass ρf.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Review is based on the extracted full text including the supplementary material; the extraction appears complete (references and supplement present) with only minor PDF line-break artifacts.
- Supplementary tables 11–17 were extracted but any figures/plots and the referenced external repo contents could not be inspected.
- No access to the actual ONNX files or the anonymous repository; all quantitative judgments derive from numbers as printed in the text, including the internal inconsistencies noted.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.