PaperlayerView exact revision on arXiv

Source-linked AI summary

GA-EIRFS: A Geometry-Augmented Repeat-Factor Sampling Method for Long-Tailed LiDAR 3D Object Detection

Taufiq Ahmed, Constantino Álvarez Casado, Daniel Herrera Castro, Sasan Sharifipour, Abhishek Kumar, Miguel Bordallo López

arXiv:2609.38116v1cs.CV

TL;DR

Long-tailed 3D detection treats imbalance as a frequency problem, although LiDAR instances with similar class frequencies can provide different geometric evidence. GA-EIRFS adds a fixed class-level geometry score to a frequency-based frame repeat factor; on nuScenes, it improves mAP and NDS across four converged runs.

  • Problem

    LiDAR classes can have similar sampling weights despite substantially different point evidence, while frequency and geometry scores are close to uncorrelated across nuScenes classes.

    Source: §1 · p. 1

  • Method

    GA-EIRFS uses a fixed class-level geometry score from annotated LiDAR points to modulate the exponent of a frequency-based repeat factor.

    Source: §1 · p. 1

  • Results

    Across four paired nuScenes runs, GA-EIRFS gains averaged +0.91 pp mAP and +0.49 pp NDS, with both metrics increasing in every run.

    Source: §5 · p. 3

  • Takeaways & Limitations

    The reported gains track sampling-weight changes rather than geometry score alone, supporting an interpretation of geometry as an amplifier of frequency-driven sampling.

    Source: §6 · p. 4

    The nuScenes evidence covers two seeds per detector, and the method was not directly compared with curriculum-based augmentation under a shared protocol.

    Source: §5 · p. 4

Abstract

from arXiv · show

Long-tailed 3D object detection is treated as a class-frequency problem, but LiDAR supervision quality depends on object observability: similar frequencies can hide different geometric evidence. We introduce Geometry-Augmented Exponentially Weighted Instance-Aware Repeat Factor Sampling (GA-EIRFS), a detector-agnostic method that modulates a frequency-based repeat factor with a fixed geometry score combining point count, surface-normal entropy, and surface coverage. GA-EIRFS changes only frame-sampling probabilities, leaving the detector and inference unchanged. On nuScenes it improves mean average precision (mAP) and the nuScenes detection score (NDS) in four converged experiments with CenterPoint and PointPillars over two seeds; for CenterPoint at seed 666, mAP rises from 0.552 to 0.563 and bicycle AP from 0.306 to 0.359. Per-class gains correlate with the class sampling-weight increase (Spearman rho=0.70, p=0.025) but not with geometry score alone (rho=0.32, p=0.37), so geometry amplifies frequency-driven need. KITTI results vary across seeds, most for the rarest class. Code: https://github.com/Multimodal-Sensing-Lab/GA-EIRFS.

1. INTRODUCTION

LiDAR long-tailed detection faces both severe class imbalance and unequal geometric evidence across classes. GA-EIRFS augments frequency-based frame resampling with a fixed class-level geometry score while leaving the detector, loss, and inference unchanged.

1. INTRODUCTION

  • Bicycles average 16.9 points per box versus 98.1 for cars and 240.3 for trailers, although bicycle and trailer receive similar frequency-based sampling weights.Across ten nuScenes classes, frequency and geometry scores are close to uncorrelated (Spearman ρ = 0.16, p = 0.65).Source: §1 · p. 1
  • GA-EIRFS uses a fixed class-level geometry score from annotated LiDAR points to modulate the exponent of the E-IRFS repeat factor.The geometry contribution is controlled by one parameter and amplifies, rather than replaces, frequency-driven adjustment.Source: §1 · p. 1
  • The method combines class frequency with class-level LiDAR observability and reduces exactly to E-IRFS when the geometry term is disabled.Source: §1 · p. 1
  • The geometry score uses three measurable properties of annotated point sets, is computed once before training, and leaves the detector, loss, and inference path unchanged.Source: §1 · p. 1
  • The evaluation covers nuScenes with two detectors and two seeds, a geometry-strength sweep, class-benefit analysis, and KITTI transfer with inconsistent gains.Source: §1 · p. 1

2. RELATED WORK

Prior long-tail 3D detection methods rebalance data, alter training objectives, or modify detectors, while existing geometry-aware approaches act within models or scenes. The paper positions GA-EIRFS as inserting a fixed sensor-derived geometry score directly into a frame repeat factor.

2. RELATED WORK

  • Geometry-aware approaches guide point selection, attention, augmentation, or object difficulty, but generally act inside a detector or synthesized scene.COM selects harder copy-paste objects using model scores; Rare Example Mining selects unlabelled tracks for annotation rather than rebalancing labelled data.Source: §2 · p. 2
  • Long-tail point-cloud methods rebalance data, alter objectives, or change detectors, while GA-EIRFS targets frame sampling without changing the detector, loss, or inference.Source: §2 · p. 2
  • The paper identifies a gap in combining fixed, sensor-derived class observability with annotation-frequency-based frame sampling.Other approaches that separate rarity from difficulty estimate difficulty from model behavior or select data for annotation.Source: §2 · p. 2

3. GEOMETRY-AUGMENTED REPEAT FACTOR SAMPLING

GA-EIRFS combines training-split frequency and geometry priors to produce frame-sampling probabilities, changing the training loop at the sampler. Its geometry score captures point support, surface-normal diversity, and observed surface coverage.

3.1. Frequency prior

  • The frequency prior increases as frame or box frequency falls, treating classes with comparable occurrence statistics identically.It follows E-IRFS and uses a threshold to determine when oversampling becomes active.Source: §3.1 · p. 2

3.2. Class-level geometry prior

  • The point-support component assigns greater geometric difficulty to classes with fewer mean in-box LiDAR points, using a log-scaled count normalized across classes.In nuScenes, traffic cones score 1.00 with 9.2 points per instance, while buses score 0.00 with 312.4.Source: §3.2 · p. 2
  • The surface-diversity component measures class-level entropy over local surface-normal orientations, excluding instances with fewer than ten points because their normal estimates are unstable.Assigning those instances zero entropy would artificially lower estimated diversity for sparsely sampled classes.Source: §3.2 · p. 2
  • The coverage component measures sparse observed surface support using box surface area rather than volume, which would make large hollow boxes appear artificially empty.The geometry score is a convex combination of the density, entropy, and coverage components.Source: §3.2 · p. 2

3.3. Geometry-augmented repeat factor

  • The class repeat factor combines frequency and geometry, recovers E-IRFS when β = 0, and determines each frame’s sampling weight through the maximum class factor in that frame.Geometry scales a frequency-driven adjustment rather than promoting geometrically difficult common classes on their own.Source: §3.2 · p. 2, §3.3 · p. 3

4. EXPERIMENTAL SETUP

The evaluation pairs baseline and GA-EIRFS sampling under matched detector, schedule, augmentation, and seed settings. It tests nuScenes with two detectors and KITTI transfer with PointPillars.

4. EXPERIMENTAL SETUP

  • The nuScenes evaluation uses CenterPoint and PointPillars over two seeds, with paired baseline-to-GA-EIRFS comparisons differing only in the sampler.The principal runs use 20 epochs; the geometry-strength sweep uses 12 epochs.Source: §4 · p. 3
  • KITTI transfer uses PointPillars across three seeds, with unchanged α, t, and β and a geometry score recomputed from KITTI point clouds.KITTI uses a 64-beam sensor, single sweeps, and three classes; its Car-to-Cyclist ratio is about 19.6:1 versus 41.5:1 on nuScenes.Source: §4 · p. 3
  • The study reports nuScenes mAP, NDS, and per-class AP, plus KITTI moderate-difficulty 3D AP measured via box IoU.NDS combines mAP with true-positive errors in translation, scale, orientation, velocity, and attribute.Source: §4 · p. 3

5. RESULTS AND DISCUSSION

GA-EIRFS improves both nuScenes metrics across four converged runs, with gains associated more strongly with sampling-weight changes than geometry score alone. Class-level outcomes and KITTI transfer are less consistent, and the authors bound their conclusions by the limited evaluation scope.

Overall performance

  • E-IRFS matched vanilla mAP at 0.5303 and 0.5361 across two seeds, while adding geometry raised it to 0.5398 and 0.5428.These 12-epoch comparisons isolate the geometry term from exponential frequency weighting.Source: §5 · p. 3
  • +0.91 pp mAP and +0.49 pp NDS were the mean gains across four paired nuScenes runs; both 95% confidence intervals excluded zero.GA-EIRFS increased both metrics in every run, with gains of 0.4–1.3 pp mAP and 0.3–0.7 pp NDS.Source: Table 2 · p. 3, §5 · p. 3

Which classes benefit

  • For CenterPoint at seed 666, bicycle AP rose from 0.306 to 0.359, while mean AP for the five rarest classes increased from 0.393 to 0.410.Across four runs, bicycle and motorcycle improved in all four; trailer split two wins and two losses.Source: §5 · p. 3

Why they benefit

  • Class sampling-weight gain correlated with AP change (Spearman ρ = 0.70, p = 0.025), unlike geometry score alone (ρ = 0.32, p = 0.37).The authors note that class sampling-weight change is not observed exposure, because frame weights depend on the maximum class weight and are normalized across frames.Source: §5 · p. 3

Qualitative behaviour

  • In validation sample 3896, GA-EIRFS recovered both sparsely sampled bicycles at scores 0.33 and 0.52, while vanilla and E-IRFS returned none above threshold.The panels differ only in the sampler; the authors caution that one frame does not establish effect size.Source: Figure 4 · p. 4, §5 · p. 4

Geometry-strength sweep

  • In Figure 3(b), bicycle AP rose monotonically from 0.283 at β = 0 to 0.354 at β = 2, while overall mAP stayed within 0.34 pp for every nonzero β.Source: §5 · p. 4

Transfer to KITTI

  • KITTI gains were inconsistent across three seeds: Pedestrian averaged +0.69 pp, Cyclist −0.16 pp, and Car was unchanged.The authors treat KITTI as a boundary condition, and note that the nuScenes evidence covers two detectors and two seeds; the class-level score cannot represent range, weather, or occlusion variation.Source: §5 · p. 4

6. CONCLUSION

The conclusion presents GA-EIRFS as a fixed class-level geometry prior added to a frequency-based frame repeat factor. NuScenes results are consistently positive across four detector-seed combinations, while KITTI transfer is inconsistent.

Conclusion

  • GA-EIRFS inserts a fixed class-level LiDAR geometry prior into a frequency-based frame repeat factor.The conclusion characterizes its gains as tracking sampling-weight change rather than geometry score alone.Source: §6 · p. 4
  • On nuScenes, mAP and NDS increased in all four converged detector-seed combinations, and CenterPoint bicycle AP rose by 5.3 pp at seed 666.Bicycle AP grew monotonically with β.Source: §6 · p. 4
  • The evidence is limited to two seeds per detector and one dataset because KITTI transfer was inconsistent; learning mixture weights and conditioning geometry on range or instance are proposed next steps.The authors interpret the method as amplifying frequency-driven sampling rather than independently sampling difficulty.Source: §6 · p. 4
Loading 2609.38116v1…