Research map / Detection and segmentation / 3D detection and driving
BEVNeXt: Reviving Dense BEV Frameworks for 3D Object Detection
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 BEVNeXt, a modernized dense BEV framework for camera-based 3D object detection that adds CRF-modulated depth estimation, Res2Fusion long-term temporal aggregation, and a two-stage object decoder with perspective refinement, reaching 64.2 NDS on nuScenes test.
Problem
Dense BEV-based camera 3D object detectors lag behind query-based (sparse) methods due to weak 2D depth modeling, limited long-term temporal receptive fields, and feature distortion during view transformation.
Approach
Modernize the BEVPoolv2 dense BEV pipeline with three additions: CRF color-smoothness modulation of predicted depth probabilities, Res2Fusion (Res2Net-style multi-scale grouped temporal fusion that avoids ego-motion warping), and a CenterPoint-style object decoder that refines object BEV features with depth-embedded spatial cross-attention.
Key results
- Achieves 64.2 NDS on nuScenes test set with V2-99 backbone (vs. 63.8 for Sparse4Dv2), and 62.2 NDS on val with ViT-Adapter-L.
- With ResNet50 at 256x704, reaches 54.8 NDS on val, outperforming SOLOFusion (53.4) and SparseBEV (54.5).
- Ablation: integrated components bring +2.2 NDS and +3.1 mAP over the BEVPoolv2 baseline; CRF gives +1.9 NDS under sparse supervision.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BEVNeXt instead of SparseBEV* | ResNet50 camera 3D detector · nuScenes val | NDS: 0.558 -> 0.548 | Hurt |
| Use BEVNeXt instead of StreamPETR | ViT-L / ViT-Adapter-L camera 3D detector · nuScenes val | NDS: 0.609 -> 0.622 | Helped |
| Use BEVNeXt instead of Sparse4Dv2 | V2-99 camera 3D detector · nuScenes test | NDS: 0.638 -> 0.642 | Helped |
| Add CRF modulation (ablation) | BEVPoolv2 with ResNet50, 256 × 704 input · nuScenes val | NDS: 0.490 -> 0.492 | Helped |
| Add CRF modulation (ablation) | BEVPoolv2 with ResNet50, 384 × 1056 input · nuScenes val | NDS: 0.502 -> 0.521 | Helped |
| Add CRF modulation (ablation) | BEVPoolv2 with ResNet101, 512 × 1408 input · nuScenes val | NDS: 0.535 -> 0.553 | Helped |
| Add Res2Fusion (ablation) | BEVPoolv2 baseline · nuScenes val | NDS: 0.526 -> 0.537 | Helped |
| Use F1/8 instead of F1/16 depth features (ablation) | BEVPoolv2 with Res2Fusion · nuScenes val | NDS: 0.537 -> 0.540 | Helped |
| Add CRF modulation to the F1/8 depth variant (ablation) | BEVPoolv2 with Res2Fusion · nuScenes val | NDS: 0.540 -> 0.542 | Helped |
| Add perspective refinement (ablation) | BEVPoolv2 with Res2Fusion, F1/8 depth, and CRF · nuScenes val | NDS: 0.542 -> 0.548 | Helped |
| Apply ego-motion transformation to historical BEVs (ablation) | Res2Fusion with window size 3 · nuScenes val | NDS: 0.537 -> 0.535 | Hurt |
| Use window size 4 instead of 3 (ablation) | Res2Fusion without ego-motion transformation · nuScenes val | NDS: 0.537 -> 0.529 | Hurt |
Surprising findings
- CRF modulation has only a minor effect (+0.2 NDS) when dense LiDAR supervision is available, but a larger effect under sparse supervision.
- Forcing ego-motion warping of historical BEV features degrades performance relative to the warping-free Res2Fusion.
Limitations
- Still less efficient than query-based methods such as StreamPETR.
- Integrating BEV frameworks into long-range settings presents challenges.
- Gains from CRF and long-term fusion saturate in long-term temporal settings.
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes val nuScenes test
Metrics: NDS mAP mATE mASE mAOE mAVE mAAE AMOTA AMOTP IDS
Baselines: BEVDepth BEVPoolv2 SOLOFusion VideoBEV Sparse4Dv2 StreamPETR SparseBEV Far3D FB-BEV BEVFormerv2
Builds on: BEVPoolv2 BEVDepth SOLOFusion CenterPoint BEVFormer
Closest papers by meaning
- RaCFormer: Towards High-Quality 3D Object Detection via Query-based Radar-Camera Fusion (CVPR 2025 · similarity 0.88)
- GeoBEV: Learning Geometric BEV Representation for Multi-view 3D Object Detection (AAAI 2025 · similarity 0.88)
- Instance-Aware Multi-Camera 3D Object Detection with Structural Priors Mining and Self-Boosting Learning (AAAI 2024 · similarity 0.87)
- UniMODE: Unified Monocular 3D Object Detection (CVPR 2024 · similarity 0.86)
- BEVDiffuser: Plug-and-Play Diffusion Model for BEV Denoising with Ground-Truth Guidance (CVPR 2025 · similarity 0.86)
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 2025 · similarity 0.86)
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.