Research map / Detection and segmentation / 3D detection and driving
Far3D: Expanding the Horizon for Surround-View 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.
Far3D extends surround-view camera 3D object detection to long perception ranges (up to 150m) using adaptive queries, perspective-aware aggregation, and 3D denoising on top of a sparse query-based framework.
Problem
Surround-view 3D object detection from cameras has focused on close ranges; extending detection to long distances causes heavy computation, unstable convergence, poor 3D recall, and query error propagation.
Approach
Combines adaptive query generation from 2D detections, perspective-aware (camera-aware SE-encoded) feature aggregation with deformable attention, and 3D denoising of GT-perturbed queries, built on StreamPETR-style propagated queries, to scale perception range to 150m.
Key results
- On Argoverse2 with 150m range, Far3D improves mAP by about 4% over prior methods, per the rebuttal.
- Trained at 150m, Far3D reaches 9.9 mAP in 50-150m versus 7.4 for StreamPETR, while also improving 0-50m mAP over StreamPETR trained at 50m (38.7 vs 34.3).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Far3D instead of StreamPETR, trained at 150m | StreamPETR · Argoverse2 50-150m | mAP: 7.4 -> 9.9 | Helped |
| Use Far3D instead of StreamPETR, trained at 150m | StreamPETR · Argoverse2 50-150m | Recall: 17.2 -> 20.5 | Helped |
| Use Far3D instead of StreamPETR, trained at 150m | StreamPETR · Argoverse2 0-50m | mAP: 31.8(-2.5) -> 37.7(-1.0) | Helped |
| Use Far3D instead of StreamPETR, trained at 150m | StreamPETR · Argoverse2 0-50m | Recall: 46.9(-5.0) -> 51.9(-3.1) | Helped |
| Train StreamPETR at 150m instead of 50m | StreamPETR · Argoverse2 0-50m | mAP: 34.3 -> 31.8(-2.5) | Hurt |
| Train StreamPETR at 150m instead of 50m | StreamPETR · Argoverse2 50-150m | mAP: 1.6 -> 7.4(+5.8) | Helped |
| Train Far3D at 150m instead of 50m | Far3D · Argoverse2 0-50m | mAP: 38.7 -> 37.7(-1.0) | Hurt |
| Train Far3D at 150m instead of 50m | Far3D · Argoverse2 50-150m | mAP: 1.9 -> 9.9(+8.0) | Helped |
Surprising findings
- Training StreamPETR with a 150m range degrades close-range (0-50m) mAP by 2.5 points, a degradation Far3D largely mitigates.
Limitations
- Long-range mAP remains low on Argoverse2 due to long-tail categories and strict long-range criteria.
- Reviewers noted limited detail on temporal modeling and training specifics in the paper; code release is promised rather than given.
Experimental setup
Datasets: nuScenes Argoverse2
Benchmarks: nuScenes Argoverse2
Metrics: mAP recall mATE
Baselines: StreamPETR PETR BEVFormer Sparse4D DETR3D
Builds on: StreamPETR PETR DETR3D BEVDepth Sparse4D
Closest papers by meaning
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 2025 · similarity 0.88)
- MonoDGP: Monocular 3D Object Detection with Decoupled-Query and Geometry-Error Priors (CVPR 2025 · similarity 0.87)
- VSRD: Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection (CVPR 2024 · similarity 0.87)
- Sparse3D: Distilling Multiview-Consistent Diffusion for Object Reconstruction from Sparse Views (AAAI 2024 · similarity 0.87)
- SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object Detection (CVPR 2024 · similarity 0.86)
- Improving Distant 3D Object Detection Using 2D Box Supervision (CVPR 2024 · 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.