A transformer that learns to look down at the road

What it does

PersFormer is an end-to-end monocular 3D lane detector. It generates bird’s-eye-view features by attending to front-view image regions via a Transformer module that uses camera parameters as a reference, then decodes 3D lane geometry. It simultaneously detects 2D lanes through an auxiliary branch that shares anchors with the 3D task, which the authors say enhances feature consistency.

The interesting bit

Instead of relying on LiDAR or stereo rigs, the model learns to warp perspective into top-down space through attention. A unified 2D/3D anchor design and the auxiliary 2D task keep the backbone features consistent between the two views.

Key highlights

• Scores F-Scores of 50.5 on OpenLane v1.1 and 53.1 on v1.2, beating GenLaneNet and 3DLaneNet in the same table.
• Achieves 72.07% F1 with 0.086 m chamfer distance on the ONCE_3DLanes benchmark.
• The 2D branch is explicitly based on LaneATT, while the overall repo draws code from Gen-LaneNet and Deformable DETR.
• Presented at ECCV 2022 as an Oral (2.7% acceptance rate).

Caveats

• Needs at least one GPU to run; CPU-only machines need not apply.
• The evaluation pipeline had a bug fix in November 2022, so older commits may not reproduce the current numbers.
• On OpenLane 2D detection, it trails dedicated 2D methods like CondLaneNet-L (42.0 vs 59.1 F-Score), confirming the 2D branch is auxiliary, not the main attraction.

Verdict

Grab it if you need a strong monocular 3D lane baseline for OpenLane or ONCE. Look elsewhere for pure 2D lane detection or lightweight CPU inference.