Research map / Detection and segmentation / 3D detection and driving
Improving Bird's Eye View Semantic Segmentation by Task Decomposition
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.
Decomposes monocular BEV segmentation into a BEV autoencoder trained to reconstruct noisy-latent BEV maps (polar coordinates) and a column-wise transformer that maps RGB features into that frozen latent space.
Problem
Direct end-to-end prediction of BEV segmentation maps from monocular perspective RGB images is hard to optimize because input and target views differ, producing distorted layouts.
Approach
Split the task into two stages: first learn a BEV latent space via a noise-corrupted autoencoder that captures typical traffic layout priors; second align RGB features to that latent space with a column-wise transformer, after converting BEV maps from Cartesian to polar coordinates so that each column corresponds to an image column.
Key results
- On nuScenes, TaDe reaches 29.2 mIoU, above GitNet (27.9) and PON (19.1) on the validation set.
- On Argoverse, TaDe reaches 22.3 mIoU versus 17.0 for PON.
- TaDe runs at 51.4 FPS with 48.5G FLOPs versus 43.8 FPS for PON, and needs 80 GPU hours versus 160 for TIM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TaDe instead of GitNet | monocular BEV segmentation · nuScenes validation | mean IoU: 27.9 -> 29.2 | Helped |
| use TaDe instead of PON | monocular BEV segmentation · Argoverse validation | mean IoU: 17.0 -> 22.3 | Helped |
| add coordinate system transformation from Cartesian to polar coordinates (ablation) | TaDe BEV segmentation on nuScenes · nuScenes validation | mean IoU: 21.6 -> 24.4 | Helped |
| add task decomposition with a BEV autoencoder (ablation) | TaDe BEV segmentation on nuScenes · nuScenes validation | mean IoU: 26.7 -> 28.7 | Helped |
| add the column-wise transformer (ablation) | TaDe BEV segmentation on nuScenes · nuScenes validation | small-object IoU: 13.9 -> 15.9 | Helped |
| fine-tune the BEV decoder (ablation) | TaDe BEV segmentation on nuScenes · nuScenes validation | mean IoU: 28.7 -> 29.2 | Helped |
Limitations
- Relies on multi-stage training with a frozen autoencoder and separate fine-tuning stage.
- Evaluated only on nuScenes and Argoverse with monocular/surround-camera setups from prior work.
Experimental setup
Datasets: nuScenes Argoverse
Benchmarks: nuScenes Argoverse
Metrics: IoU mIoU FPS FLOPs GPU hours
Baselines: IPM VED VPN Sim2real OFT PON EPOSH DiffBEV GitNet TIM
Builds on: PON TIM GitNet CVT
Closest papers by meaning
- A Black-Box Evaluation Framework for Semantic Robustness in Bird’s Eye View Detection (AAAI 2025 · similarity 0.89)
- Predictive Uncertainty Quantification for Bird's Eye View Segmentation: A Benchmark and Novel Loss Function (ICLR 2025 · similarity 0.88)
- VQ-Map: Bird's-Eye-View Map Layout Estimation in Tokenized Discrete Space via Vector Quantization (NeurIPS 2024 · similarity 0.88)
- M-BEV: Masked BEV Perception for Robust Autonomous Driving (AAAI 2024 · similarity 0.88)
- MamBEV: Enabling State Space Models to Learn Birds-Eye-View Representations (ICLR 2025 · similarity 0.87)
- Map It Anywhere: Empowering BEV Map Prediction using Large-scale Public Datasets (NeurIPS 2024 · similarity 0.87)
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.