acceptodds
Under review as a conference paper at ICLR 2027

Beyond Geometric Augmentation: Hybrid Equivariant Lidar Detection

Abstract

LiDAR detectors rely heavily on geometric augmentation to learn how object states change with scene geometry. We argue that the burden of geometric ro- bustness should not fall on augmentation alone: geometric structure should also be built into the detector. We investigate this principle through HybridFusion, which couples a steerable equivariant pathway with a non-equivariant stream. Field-wise gate conditioning connects the streams throughout the network, while cross-stream attention and residual merging combine their features in the detec- tion head. On nuScenes, a detailed TransFusion-L study compares standard, pre- dominantly equivariant, and hybrid detectors with and without global geometric augmentation, with balanced resampling disabled. HybridFusion improves mAP by 6.55 percentage points without global augmentation and by 4.83 points with augmented training, with large reductions in orientation and velocity errors. Aug- mentation remains valuable: it adds a further 9.27 mAP points to the hybrid model. Improvements in mAP and NDS also extend to augmented CenterPoint and SEC- OND. A non-equivariant two-stream control fails to recover the hybrid gains, and removing gate conditioning lowers both scores. These detection results support sharing the burden between architecture and augmentation. Hybrid equivariance builds transformation structure into feature processing, improving prediction be- yond what the augmented standard detector achieves while retaining the benefits of augmented training.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Autonomous Driving, Equivariant Machine Learning, Representation Learning, Geometric Data Augmentation

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

This paper studies the interaction between geometric augmentation and architectural equivariance for LiDAR detection. It proposes HybridFusion, which combines an equivariant pathway with a conventional non-equivariant stream through field-wise conditioning and feature fusion. Experiments on nuScenes show improvements over standard TransFusion-L with and without geometric augmentation, with additional results on CenterPoint and SECOND.

Critical strengths
  • The paper investigates an interesting relationship between geometric augmentation and architectural inductive bias. In particular, the hybrid design recognizes that strict equivariance may be overly restrictive for real driving scenes, while relying entirely on augmentation requires geometric behavior to be learned from data.

  • The reported improvements on TransFusion-L are substantial. HybridFusion improves both mAP and NDS, with particularly notable reductions in orientation and velocity errors.

  • The architectural controls are useful. EquiFusion underperforms the standard detector, while the non-equivariant two-stream control does not reproduce the HybridFusion gains. The FiLM ablation further supports the contribution of the proposed interaction between the two streams.

Critical weaknesses
  1. The experiments do not directly establish whether HybridFusion achieves the intended balance between equivariant structure and non-equivariant flexibility.

A central motivation for HybridFusion is that strict equivariance may be too restrictive, while a conventional detector must learn geometric behavior primarily from augmented examples. HybridFusion therefore combines a structured equivariant pathway with an unconstrained stream. However, the latter explicitly breaks exact equivariance through FiLM conditioning and head-level fusion. The authors consequently describe equivariance as a structured feature-processing bias rather than an end-to-end property.

The current experiments show that HybridFusion achieves higher detection accuracy than either TransFusion-L or EquiFusion, but they do not characterize how much geometric consistency/equivariance the hybrid model actually retains. Therefore, it is difficult to determine whether the improvement reflects the intended compromise between geometric structure and flexibility, or simply a more effective architecture. A controlled evaluation under transformed inputs, or a direct measure of equivariance error, would help establish this central interpretation and clarify how Standard, EquiFusion, and HybridFusion differ in the tradeoff between geometric consistency and detection accuracy.

  1. It is unclear how much of the reported advantage remains under a strong, matched training recipe.

The core experiments disable balanced resampling, under which augmented HybridFusion improves TransFusion-L by 4.83 mAP and 9.92 NDS points. However, with balanced resampling, HybridFusion achieves 0.6548/0.7024 mAP/NDS compared with the published OpenPCDet baseline of 0.6458/0.6943, reducing the difference to only 0.90/0.81 points. Moreover, this is not a controlled comparison because the standard result is externally reported rather than reproduced using the same experimental setup. The authors explicitly acknowledge this limitation.

This raises uncertainty about the practical magnitude of the architectural contribution under a stronger conventional training pipeline. A matched standard-versus-HybridFusion comparison using the same balanced-resampling and tuned training recipe would be important for determining whether the substantial gains reported in the main experiments persist under stronger training conditions.

Questions for authors
  1. Can the authors characterize the tradeoff between equivariance and flexibility more directly? For example, evaluating Standard, EquiFusion, and HybridFusion under controlled test-time rotations or reporting an equivariance-consistency metric could reveal how much geometric structure HybridFusion retains and whether this correlates with its detection improvements.

  2. Can the authors provide a matched comparison under balanced resampling and a fully tuned training recipe? This seems particularly important given that the advantage over the published OpenPCDet baseline is substantially smaller than in the main experiments.

  3. How stable are the reported improvements across training runs? Each configuration currently uses only one seed. Multiple runs would be especially helpful for the architectural controls and for CenterPoint/SECOND, where the mAP improvements are only around 1.4 points.

  4. What is the computational overhead of HybridFusion? The paper reports approximately 7–8% more parameters, but runtime, memory consumption, or FLOPs would better characterize the practical cost of the additional stream and steerable processing.

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

The paper examines the relationship between equivariant feature processing and geometric augmentation in LiDAR-based 3D detection. EquiFusion uses sparse SO(3) steerable layers, followed by restriction and projection to SO(2) BEV features. HybridFusion adds a non-equivariant branch. The branches interact through field-wise scalar gates, cross-attention, and residual merging. The authors evaluate three TransFusion-L variants on nuScenes, with and without global augmentation. Additional experiments cover CenterPoint, SECOND, a non-equivariant two-stream control, and a no-FiLM ablation. The reported results favor HybridFusion, while augmentation provides further gains. The authors state that the full detector is not strictly equivariant.

Critical strengths
  1. The model combines equivariant and non-equivariant features in a sparse-to-BEV detector. Field-wise scalar gating preserves the coupling between field components while allowing modulation by the non-equivariant branch. This provides a reasonable mechanism for combining geometric structure with flexible feature processing. The authors also clearly describe the limits of the model's equivariance.

  2. The reported improvements extend across three detectors. With augmentation, TransFusion-L improves by 4.83 mAP and 9.92 NDS percentage points, with gains also reported for CenterPoint and SECOND. Removing FiLM under the same training settings reduces mAP by 1.49 points and NDS by 1.57 points. This ablation provides evidence for the contribution of FiLM within the proposed model.

Critical weaknesses
  1. Differences in training settings make it hard to obtain a clear attribution of the gains to the proposed architecture. Supplementary Tables 7 and 8 show different learning rates and flipping policies for TransFusion-L. SECOND also differs in rotation range, scale range, and translation augmentation. The ground-truth sampling schedules differ for CenterPoint and SECOND. These comparisons therefore do not isolate the effect of equivariant processing or its interaction with augmentation. The two-stream control also has fewer parameters than HybridFusion: 8.409M versus 8.974M. Under balanced resampling, the reported gains are 0.90 mAP and 0.81 NDS points. However, this comparison uses an external baseline that was not reproduced in the same implementation. Whether the gains persist under matched training conditions remains unresolved and requires further experiments.

  2. The contribution beyond prior work is not sufficiently established. TED studies equivariant detection, while Residual Pathway Priors combines equivariant and non-equivariant processing. The paper acknowledges both precedents. Its potential novelty lies in the specific sparse-to-BEV design and the interactions between branches. However, there is no comparison with a simple residual hybrid under matched training conditions. The FiLM ablation shows a benefit within this architecture, but does not establish an advantage over simpler gating alternatives. The paper also lacks evaluation on rotated inputs and measurements of feature or output equivariance. These tests are directly relevant to the claim of geometric robustness. In the augmented TransFusion-L comparison, approximately 5.95 of the 9.92 NDS-point gain comes from the velocity term. This is calculated using the official NDS formula, including clipping, and the rounded values in Table 1. Translation and scale errors increase slightly. The source of the large velocity improvement remains insufficiently explained. The authors acknowledge the absence of strict equivariance. My concern is that the evidence does not yet link the reported gains to the proposed geometric structure.

Questions for authors
  1. What gains remain when standard TransFusion-L and HybridFusion use identical augmentation and ground-truth sampling settings, with comparable learning-rate tuning budgets? Could authors reproduce the balanced-resampling baseline in the same implementation, with results reported across multiple training seeds?

  2. What explains the reduction in mAVE from 1.1424 to 0.4053 under augmented training? Could authors provide a per-class breakdown and clarify whether velocity loss, decoding, and velocity-label transformations are identical across the two models.

  3. Does HybridFusion produce more stable predictions than the baseline under test-time input rotations? What equivariance error is measured in the structured branch before and after gate conditioning?

  4. How does the method compare with a simple residual combination of equivariant and non-equivariant features under matched training settings and comparable parameter counts? What evidence supports field-wise gating over simpler alternatives?

Reviewer_R5HWOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

HybridFusion combines a steerable equivariant pathway with an unconstrained stream through field-wise gating and head-level fusion. On nuScenes, it improves detection with and without geometric augmentation, with additional results on CenterPoint and SECOND.

Critical strengths
  1. I like the motivation for combining geometric structure with flexibility for driving-scene asymmetries. The six-configuration TransFusion-L study shows that both architecture and augmentation contribute to detection performance.

  2. The non-equivariant two-stream and no-FiLM controls help test the architectural choices. Improvements on CenterPoint and SECOND also suggest that the approach is useful beyond one detection head.

Critical weaknesses
  1. The standard and hybrid training recipes are not matched: Appendix B changes TransFusion-L's learning rate from 0.002 to 0.004, disables hybrid flips, and changes other detectors' augmentation or sampling schedules. These differences confound how much of the gain comes from architecture.

  2. The geometric-robustness claim is supported by clean nuScenes scores without transformed-input testing, and each configuration uses one training seed. This leaves uncertainty about symmetry-specific benefits and the smaller gains under balanced resampling.

Questions for authors
  • Do the gains persist with matched learning-rate searches, augmentation policies, and ground-truth sampling schedules?

  • How do the models compare under controlled test-time yaw rotations, and across multiple seeds?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.