Camera-based bird's-eye view (BEV) perception methods such as Fast-BEV have demonstrated promising results with fast inference speed and competitive accuracy. However, their robustness to real-world corruptions remains insufficiently explored, posing risks for autonomous driving deployment where environmental variations are inevitable. In this paper, we propose Sparse Recovery Bird's-Eye View Model Soup (SR-BEVS), a sparse recovery learning approach that simultaneously improves clean accuracy and corruption robustness without requiring additional data or architectural modifications. Our method identifies pruning candidates based on layer-wise weight magnitudes, applies random pruning to create diverse sparse models, and retrains each model with a randomly sampled learning rate. By averaging these diverse recovered models into a single dense model, we achieve enhanced robustness against various corruptions while introducing zero inference overhead. We further provide a mode connectivity analysis demonstrating that our pruning-based diversification preserves the connected low-loss basin necessary for stable weight averaging. Experiments on the nuScenes dataset demonstrate up to 33.8% improvement in corruption robustness and a 1.8% gain in clean accuracy over the Fast-BEV baseline, with consistent improvements across sensor noise, weather distortions, and illumination changes. Our simple yet effective approach may be applicable to other BEV models.
Figures & tables
Fig. 1 : The mAP (a) and NDS (b) performance improvement rates (%) of Model Soup and SR-BEVS compared to the Fast-BEV baseline on the M0 model [ 1 ] . Results are averaged across severity levels for each of the seven corruption types, plus the nuScenes clean data [ 2 ] . The results show that our SR-BEVS consistently outperforms the two baselines.
Fig. 2 : Qualitative comparison under clean and corrupted inputs. Fast-BEV performs well on clean data but produces incorrect predictions under heavy fog, whereas SR-BEVS maintains accurate detection even under severe corruption, yielding more stable and spatially consistent bounding boxes across the scene.
Fig. 3 : Weight diversity comparison between Model Soup and SR-BEVS. (a) L2 distance (higher is better) and (b) cosine similarity (lower is better) between candidate model pairs. Standard Model Soup candidates exhibit high similarity, while SR-BEVS produces substantially more diverse models.
Fig. 4 : Linear interpolation between SR-BEVS candidate model pairs on the M0 configuration. The horizontal axis denotes the interpolation coefficient α in θ(α)=(1−α)θa+αθb . Both mAP (left) and NDS (right) remain stable across the entire path, indicating the absence of a loss barrier between candidates.
Fig. 5 : 2D loss landscape visualization on the M0 configuration. Candidate models are projected onto the two principal directions of the weight space. (a) Standard Model Soup candidates cluster tightly near the pretrained model. (b) SR-BEVS candidates are more spread out while remaining within the same low-loss basin, demonstrating increased diversity without breaking mode connectivity.
Method
Gauss Noise
Gauss Blur
Shot Noise
Fog
Contr.
Bright.
Occl.
5%
10%
15%
σ 1
σ 2
σ 3
5%
10%
M
H
L
H
5%
10%
20%
10%
20%
M0
mAP
Fast-BEV
0.189
0.079
0.017
0.227
0.125
0.051
0.221
0.219
0.234
0.173
0.266
0.197
0.276
0.276
0.275
0.226
0.180
Fast-BEV (FT)
0.190
0.081
0.017
0.227
0.126
0.052
0.222
0.219
0.234
0.174
0.267
0.198
0.276
0.277
0.275
0.227
0.180
Model Soup
0.190
0.080
0.016
0.229
0.127
0.053
0.222
0.220
0.236
0.175
0.266
0.198
0.278
0.278
0.277
0.228
0.182
SR-BEVS
0.198
0.087
0.018
0.231
0.129
0.056
0.230
0.228
0.247
0.192
0.270
0.202
0.282
0.281
0.281
0.233
0.187
NDS
Fast-BEV
0.325
0.181
0.077
0.373
0.290
0.153
0.360
0.359
0.378
0.325
0.403
0.339
0.412
0.411
0.409
0.369
0.324
TABLE I : Robustness evaluation on nuScenes validation set under diverse corruptions and severity levels. Fast-BEV (FT) uses the same additional fine-tuning budget as SR-BEVS recovery.
Method
M0 (R18)
M2 (R34)
mAP ↑
NDS ↑
mAP ↑
NDS ↑
Fast-BEV
0.277
0.411
0.326
0.455
Fast-BEV (FT)
0.278
0.412
0.327
0.457
Model Soup
0.278
0.412
0.330
0.457
SR-BEVS
0.282
0.417
0.332
0.460
TABLE II : Detection performance on nuScenes validation set. Fast-BEV (FT) denotes additional fine-tuning with the same training budget as SR-BEVS recovery.
Fig. 6 : Qualitative comparison of 3D detection results under severe corruptions on the M0 configuration. Top row: under strong Gaussian noise, Fast-BEV and Model Soup misclassify noise patterns as objects such as pedestrians (green boxes), while SR-BEVS avoids these incorrect predictions. Bottom row: under heavy fog, Fast-BEV and Model Soup miss objects present in the scene, whereas SR-BEVS maintains reliable detection consistent with the ground truth.
Pruning Candidate Ratio ( p )
mAP ↑
NDS ↑
Baseline
0.277
0.411
p=0.1
0.280
0.414
p=0.2
0.282
0.417
p=0.3
0.279
0.413
TABLE III : Effect of pruning candidate ratio p on model performance (M0, clean data)
#Candidates ( N )
mAP ↑
NDS ↑
1
0.278
0.412
2
0.280
0.414
3
0.282
0.417
5
0.284
0.419
TABLE IV : Effect of the number of candidate models N on model performance (M0, clean data)
Model
Method
Weight Diversity
Clean
Corrupted (Avg.)
Cos.Sim. ↓
L2 Dist. ↑
mAP
NDS
mAP
NDS
M0
Soup
0.914
1.706
0.278
0.412
0.191
0.324
SR-BEVS
0.912
5.064
0.282
0.417
0.197
0.333
M2
Soup
0.900
5.795
0.330
0.457
0.233
0.373
SR-BEVS
0.893
24.389
0.332
0.460
0.241
0.381
TABLE V : Weight Diversity Analysis and Performance Comparison
Corruption robustness studies how to keep learning systems reliable when data, inputs, labels, or feedback are noisy, missing, or adversarially altered.
Stuart Weitzman School of Design University of Pennsylvania Philadelphia, USA · Courant Institute of Mathematical Science New York University New York, USA