Research map / Detection and segmentation / 3D detection and driving
SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects
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 SeaBird, a sequential BEV foreground segmentation pipeline trained with dice loss, plus a theoretical argument that dice loss is more noise-robust than regression losses for large objects, improving monocular 3D detection of large objects on KITTI-360 and nuScenes.
Problem
Monocular 3D detectors perform well on cars but poorly on large objects (trucks, buses, trailers), even on nearly balanced datasets; the paper attributes this partly to noise sensitivity of depth regression losses.
Approach
Train a BEV foreground segmentation head with dice loss first (which the authors prove has lower gradient variance than L1/L2 for large objects under depth noise), then fine-tune a 3D box head that takes the segmentation map concatenated with BEV features in a sequential, two-stage pipeline.
Key results
- On KITTI-360 Test, PBEV+SeaBird reaches 4.64 AP3D50 mAP and 37.12 AP3D25 mAP, outperforming all listed monocular baselines and the LiDAR baselines.
- On nuScenes Val at 640x1600 with V2-99, HoP+SeaBird improves APLrg from 36.5 to 40.3 and mAP from 49.6 to 52.7 over HoP.
- On KITTI-360 Val, I2M+SeaBird (dice) raises APLrg from 4.86 (no loss) to 8.71 and AP3D25 APLrg to 35.76 vs 26.33 without dice loss.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SeaBird to HoP | HoP, V2-99 · nuScenes Test | APLrg: 37.1 -> 38.4 | Helped |
| Add SeaBird to HoP | HoP, R101 · nuScenes Test | APLrg: 36.0 -> 36.6 | Helped |
| Add SeaBird to BEVerse-T | BEVerse-T, Swin-T · nuScenes Val | APLrg: +1.0 | Helped |
| Add SeaBird to BEVerse-S | BEVerse-S, Swin-S · nuScenes Val | APLrg: +3.7 | Helped |
| Add SeaBird to HoP | HoP, V2-99 · nuScenes Val | APLrg: +3.8 | Helped |
| Add Dice loss to BEV segmentation (ablation) | I2M+SeaBird · KITTI-360 Val | AP3D50 APLrg (%): 4.86 -> 8.71 | Helped |
| Add Dice loss instead of Smooth L1 for BEV segmentation (ablation) | I2M+SeaBird · KITTI-360 Val | AP3D50 APLrg (%): 7.63 -> 8.71 | Helped |
| Add Dice loss instead of MSE for BEV segmentation (ablation) | I2M+SeaBird · KITTI-360 Val | AP3D50 APLrg (%): 7.04 -> 8.71 | Helped |
| Add Dice loss instead of cross-entropy for BEV segmentation (ablation) | I2M+SeaBird · KITTI-360 Val | AP3D50 APLrg (%): 7.06 -> 8.71 | Helped |
| Add the BEV segmentation head (ablation) | I2M+SeaBird · KITTI-360 Val | AP3D50 APLrg (%): 7.52 -> 8.71 | Helped |
| Add the detection head (ablation) | I2M+SeaBird · KITTI-360 Val | BEV Seg IoU MFor (%): 29.25 -> 31.42 | Helped |
| Restrict BEV segmentation to foreground categories (ablation) | I2M+SeaBird · KITTI-360 Val | AP3D50 APLrg (%): 1.61 -> 8.71 | Helped |
Surprising findings
- Joint training of BEV segmentation over all categories including background hurts Mono3D performance, and frontal detectors struggle on large objects even on a nearly balanced dataset.
Limitations
- Theoretical proof is for a simplified single-layer linear case with constant object length and depth as the only error source.
- Gains on overall mAP and NDS are modest on nuScenes, and some metrics (e.g., NDS for HoP at 512x1408) slightly drop.
Experimental setup
Datasets: KITTI-360 nuScenes KITTI-360 PanopticBEV BEV segmentation GT
Benchmarks: KITTI-360 Test leaderboard KITTI-360 Val nuScenes Test nuScenes Val
Metrics: AP3D50 mAP AP3D25 mAP APLrg APCar APSml NDS mAP BEV segmentation mean IoU
Baselines: GrooMeD-NMS MonoDLE GUP Net DEVIANT Cube R-CNN MonoDETR I2M PanopticBEV BEVerse HoP
Builds on: Image2Maps PanopticBEV BEVerse HoP BoxNet
Closest papers by meaning
- UniMODE: Unified Monocular 3D Object Detection (CVPR 2024 · similarity 0.88)
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 2025 · similarity 0.86)
- MonoDGP: Monocular 3D Object Detection with Decoupled-Query and Geometry-Error Priors (CVPR 2025 · similarity 0.86)
- Predictive Uncertainty Quantification for Bird's Eye View Segmentation: A Benchmark and Novel Loss Function (ICLR 2025 · similarity 0.86)
- GeoBEV: Learning Geometric BEV Representation for Multi-view 3D Object Detection (AAAI 2025 · similarity 0.86)
- A Black-Box Evaluation Framework for Semantic Robustness in Bird’s Eye View Detection (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.