Research map / Vision-language and multimodal / Cross-modal learning
AlterMOMA: Fusion Redundancy Pruning for Camera-LiDAR Fusion Models with Alternative Modality Masking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AlterMOMA, a pruning framework for camera-LiDAR fusion models that alternately masks each backbone to reactivate fusion-redundant parameters and scores parameters with the AlterEva criterion.
Problem
Pre-trained single-modal camera and LiDAR backbones loaded into camera-LiDAR fusion models develop cross-modal feature redundancy, which existing single-modal and multi-modal pruning methods fail to identify, causing large accuracy drops when pruning fusion models.
Approach
Alternately mask one modality backbone so the fusion module must reactivate the other backbone's fusion-redundant features; score parameters with AlterEva, combining a Taylor-approximated deactivated-contribution indicator (DeCI) and a reactivated-redundancy indicator (ReRI) computed from loss changes during reactivation.
Key results
- On nuScenes with BEVfusion-mit at 90% sparsity, AlterMOMA reaches 63.5 mAP versus 58.6 for the best baseline ProsPr (and 67.8 unpruned).
- On KITTI with AVOD-FPN at 90% sparsity, AlterMOMA reaches 77.4 AP-3D easy for Car versus 74.2 for ProsPr.
- On ResNet101+SECOND structured pruning at 50% sparsity, AlterMOMA gets 64.5 mAP versus 62.5 for ProsPr, with 56.7% GFLOPs reduction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AlterMOMA instead of ProsPr at 80% sparsity | BEVfusion-mit (Swin-T + VoxelNet) · nuScenes val 3D detection | mAP: 64.3% → 67.3% | Helped |
| Use AlterMOMA instead of ProsPr at 85% sparsity | BEVfusion-mit (Swin-T + VoxelNet) · nuScenes val 3D detection | mAP: 61.9% → 65.5% | Helped |
| Use AlterMOMA instead of ProsPr at 90% sparsity | BEVfusion-mit (Swin-T + VoxelNet) · nuScenes val 3D detection | mAP: 58.6% → 63.5% | Helped |
| Use AlterMOMA instead of ProsPr across 80%, 85%, and 90% sparsity | BEVfusion-pku (Swin-T + VoxelNet) · nuScenes val 3D detection | mAP | Helped |
| Use AlterMOMA instead of ProsPr at 90% sparsity | AVOD-FPN · KITTI val Car, easy | AP-3D: 74.2% → 77.4% | Helped |
| Use AlterMOMA instead of ProsPr at 90% sparsity | AVOD-FPN · KITTI val Car, easy | AP-BEV: 81.2% → 85.3% | Helped |
| Use AlterMOMA instead of ProsPr at 80% sparsity | BEVfusion-mit (Swin-T + VoxelNet) · nuScenes val BEV map segmentation | mIoU: 57.7% → 60.7% | Helped |
| Use AlterMOMA instead of ProsPr at 85% sparsity | BEVfusion-mit (Swin-T + VoxelNet) · nuScenes val BEV map segmentation | mIoU: 56.2% → 59.2% | Helped |
| Use AlterMOMA instead of ProsPr at 90% sparsity | BEVfusion-mit (Swin-T + VoxelNet) · nuScenes val BEV map segmentation | mIoU: 54.1% → 57.7% | Helped |
| Remove ReRI by setting β/α = 0, relying only on DeCI (ablation) | AlterMOMA · nuScenes val 3D detection | mAP | Hurt |
| Increase the weight of ReRI as β/α increases from zero (ablation) | AlterMOMA · nuScenes val 3D detection | mAP | Helped |
| Overweight ReRI by raising β beyond a certain threshold (ablation) | AlterMOMA · nuScenes val 3D detection | mAP | Hurt |
Surprising findings
- At 30% structured pruning, AlterMOMA slightly exceeds the unpruned model's mAP (65.3 vs 64.6), which the authors attribute to removing redundancy.
Limitations
- Only applied to camera-LiDAR perception models; extension to vision-language or other multi-modal models is left open.
- Reactivation can also reactivate contributed parameters that may be accidentally pruned when ReRI weight is large.
- Hyperparameters alpha and beta are set to 1 and the derivation relies on a small-learning-rate approximation.
Experimental setup
Datasets: nuScenes KITTI
Benchmarks: nuScenes KITTI
Metrics: mAP NDS mIoU AP-3D AP-BEV GFLOPs
Baselines: IMP SynFlow SNIP ProsPr
Builds on: IMP SynFlow SNIP ProsPr
Closest papers by meaning
- Look Before You Fuse: 2D-Guided Cross-Modal Alignment for Robust 3D Detection (CVPR 2026 · similarity 0.87)
- Adaptive-Smooth LiDAR-Camera Knowledge Distillation with Heterogeneous Fusion for Multi-View 3D Object Detection (AAAI 2026 · similarity 0.86)
- Layer-Wise Modality Decomposition for Interpretable Multimodal Sensor Fusion (NeurIPS 2025 · similarity 0.86)
- Resilient Sensor Fusion Under Adverse Sensor Failures via Multi-Modal Expert Fusion (CVPR 2025 · similarity 0.85)
- SimDistill: Simulated Multi-Modal Distillation for BEV 3D Object Detection (AAAI 2024 · similarity 0.84)
- Fusion Is Not Enough: Single Modal Attacks on Fusion Models for 3D Object Detection (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.