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.
Figures & tables
Figure 1: Rarity and geometric difficulty are close to independent across the ten nuScenes classes. Markers are classes, placed by the E-IRFS frequency term qc and the geometry score Gc of Eq. ( 5 ). Grey curves are contours of rc ( α=2.0 , β=1.0 ): vertical distance at fixed rarity is the sampling weight that geometry adds.
Figure 2: GA-EIRFS pipeline. The priors qc and Gc of Eqs. ( 1 ) and ( 5 ) are estimated once from the training split and enter the class repeat factor of Eq. ( 6 ), which becomes a frame sampling probability.
Points/
Repeat factor rc
Class
Instances
obj.
Gc
β=0
β=1
Car
339,949
98.1
0.601
1.29
1.49
Pedestrian
161,928
11.7
0.807
1.37
1.78
Barrier
107,507
62.6
0.252
1.56
1.74
Truck
65,262
209.8
0.431
1.51
1.80
Traffic cone
62,964
9.2
0.588
1.61
2.14
Table 1: nuScenes class statistics and repeat factors, ordered by instance count. Eq. ( 6 ) uses α=2 , t=0.01 : β=0 gives E-IRFS and β=1 our default.
Figure 3: Quantitative evidence on nuScenes. (a) Per-class AP change at 20 epochs, in percentage points (pp). Bars are the mean of the two seeds, vertical markers are the individual seeds, and classes are ordered by Gc , lowest at the bottom. (b) Geometry-strength sweep at 12 epochs, CenterPoint, seed 666. (c) AP change against the exposure gain Δrc=rc(β=1)−rc(β=0) , with the Spearman rank correlation over the ten classes.
Figure 4: Top-down view of nuScenes validation sample 3896, in which the two annotated bicycles are far apart and sparsely sampled. Green boxes are ground truth, dashed red boxes are predictions scoring at least 0.3. Vanilla and E-IRFS return no bicycle above the threshold, whereas GA-EIRFS recovers both, at 0.33 and 0.52. The panels differ only in the sampler.
mAP ↑
NDS ↑
Detector
Seed
V
GA
Δ
V
GA
Δ
CenterPoint
666
0.552
0.563
+1.1
0.635
0.642
+0.7
CenterPoint
1337
0.554
0.563
+0.9
0.638
0.641
+0.3
PointPillars
666
0.385
0.388
+0.4
0.536
0.538
+0.3
PointPillars
1337
0.382
0.395
+1.3
0.534
0.541
+0.7
Mean, 4 runs
+0.9
+0.5
Table 2: nuScenes validation results after 20 epochs. V is training without rebalancing and GA is GA-EIRFS. Higher is better, Δ is in percentage points computed before rounding, and the better value of each metric within a run is in bold. The last row is the mean of the four paired differences, with 95% confidence intervals [+0.27,+1.55] for mAP and [+0.12,+0.86] for NDS.
Car
Pedestrian
Cyclist
Seed
V
GA
V
GA
V
GA
666
75.88
75.87
44.19
45.34
61.90
62.29
1337
76.28
75.81
41.93
43.74
62.69
59.99
42
75.34
75.61
44.89
44.00
61.05
62.88
Mean
75.83
75.76
43.67
44.36
61.88
61.72
SD
0.47
0.14
1.55
0.86
0.82
1.53
Table 3: KITTI 3D AP (%) at moderate difficulty, PointPillars, three seeds. Better value of each pair in bold.