Research map / Architectures and efficiency / Attention and transformers
Driving on Registers
Ellington Kirby, Alexandre Boulch, Yihong Xu, Yuan Yin, Gilles Puy, Éloi Zablocki, Andrei Bursuc, Spyros Gidaris, Renaud Marlet, Florent Bartoccioni, Anh-Quan Cao, Nermin Samet and 2 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| initialize backbone with pretrained DINOv2 instead of random (ablation) | DrivoR · navval | PDMS: 70.1 -> 90.0 | Improved |
| use DINOv2 instead of ImageNet-21k pretraining (ablation) | DrivoR · navval | PDMS: 87.5 -> 90.0 | Improved |
| use camera registers instead of pooling for token compression (ablation) | DrivoR · navval | PDMS: 89.7 -> 90.0 | Improved |
| use registers instead of a transformer decoder compressor (ablation) | DrivoR · navval | PDMS: 89.3 -> 90.0 | Improved |
| compress 16k tokens to 64 register tokens (ablation) | DrivoR · navval | PDMS: 90.2 -> 90.0 | No clear effect |
| LoRA finetuning instead of frozen backbone (ablation) | DrivoR registers · navval | PDMS: 84.4 -> 90.0 | Improved |
| LoRA finetuning instead of full finetuning (ablation) | DrivoR registers · navval | PDMS: 88.4 -> 90.0 | Improved |
| keep pre-existing DINOv2 registers in addition to new ones (ablation) | DrivoR · navval | PDMS: 89.8 vs 90.0 | No clear effect |
| increase registers per camera from 16 to 32 (ablation) | DrivoR · navval | PDMS: 90.0 -> 89.8 | No clear effect |
| increase trajectory queries from 1 to 64 (ablation) | DrivoR · navval | PDMS: 80.1 -> 90.0 | Improved |
| use separate scoring and trajectory decoders instead of one branch (ablation) | DrivoR · navval | PDMS: 84.7 -> 86.8 | Improved |
| add disentanglement (stop-gradient plus trajectory re-embedding) to scorer (ablation) | DrivoR · navval | PDMS: 86.8 -> 90.0 | Improved |
Closest papers by meaning
- DriveTransformer: Unified Transformer for Scalable End-to-End Autonomous Driving (ICLR 2025 · similarity 0.87)
- FIVE-VLA: Fast and EffectIVE Closed-Loop Autonomous Driving with Recurrent Action Memory (NeurIPS 2026 · similarity 0.87)
- DualAD: Disentangling the Dynamic and Static World for End-to-End Driving (CVPR 2024 · similarity 0.85)
- ReflectDrive-2: Reinforcement-Learning-Aligned Self-Editing for Discrete Diffusion Driving (NeurIPS 2026 · similarity 0.84)
- Prune2Drive: A Plug-and-Play Framework for Accelerating Vision-Language Models in Autonomous Driving (CVPR 2026 · similarity 0.84)
- VR-Drive: Viewpoint-Robust End-to-End Driving with Feed-Forward 3D Gaussian Splatting (NeurIPS 2025 · 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.
