[Paper Review] 3D-LaneNet+: Anchor Free Lane Detection using a Semi-Local Representation

Netalee Efrat, Max Bluvstein+4 more|arXiv (Cornell University)|Nov 1, 2020

Autonomous Vehicle Technology and Safety23 references30 citations

TL;DR

3D-LaneNet+ introduces an anchor-free, semi-local tile representation for 3D lane detection that handles arbitrary topology (splits, merges, perpendicular lanes) and outperforms the original 3D-LaneNet by learning local lane segments and clustering them with a global embedding.

ABSTRACT

3D-LaneNet+ is a camera-based DNN method for anchor free 3D lane detection which is able to detect 3d lanes of any arbitrary topology such as splits, merges, as well as short and perpendicular lanes. We follow recently proposed 3D-LaneNet, and extend it to enable the detection of these previously unsupported lane topologies. Our output representation is an anchor free, semi-local tile representation that breaks down lanes into simple lane segments whose parameters can be learnt. In addition we learn, per lane instance, feature embedding that reasons for the global connectivity of locally detected segments to form full 3d lanes. This combination allows 3D-LaneNet+ to avoid using lane anchors, non-maximum suppression, and lane model fitting as in the original 3D-LaneNet. We demonstrate the efficacy of 3D-LaneNet+ using both synthetic and real world data. Results show significant improvement relative to the original 3D-LaneNet that can be attributed to better generalization to complex lane topologies, curvatures and surface geometries.

Motivation & Objective

• Motivate and enable 3D lane detection for arbitrary lane topologies beyond parallel, long lanes.
• Propose an anchor-free, semi-local tile representation that decomposes lanes into local segments.
• Learn per-tile segment parameters and a global embedding to cluster segments into full lanes.
• Remove dependence on lane anchors and polyline fitting to improve generalization to complex geometries.
• Demonstrate improved 3D lane detection on synthetic and real-world datasets.

Proposed method

• Project image features to Bird's Eye View (BEV) via dual-pathway encoder and IPM.
• Divide BEV into non-overlapping tiles and predict per-tile lane segment parameters (offset, angle, height) and a lane-intersection score.
• Represent lane angles with a hybrid classification-regression scheme over angle bins.
• Train a global embedding per tile and apply a discriminative pull-push loss to cluster tiles into lanes.
• Use mean-shift clustering on learned embeddings to form complete lanes without anchors or polyline fitting.
• During inference, convert tile-level segments to 3D points and map from BEV to camera coordinates.

Experimental results

Research questions

• RQ1 Can an anchor-free, semi-local tile representation detect 3D lanes with arbitrary topology (splits, merges, short or perpendicular lanes) in BEV?
• RQ2 Does learning per-tile lane segments plus a global embedding improve generalization to complex curvatures and road surface geometries compared to anchor-based 3D-LaneNet?
• RQ3 Is clustering through a learned embedding more effective than greedy segment stitching for forming full lane curves?
• RQ4 How does the approach perform on synthetic and real-world datasets in terms of detection MAP and geometric accuracy?

Key findings

• 3D-LaneNet+ achieves higher MAP and AP50 than the original 3D-LaneNet on synthetic data (MAP 0.90 vs 0.74; AP50 0.95 vs 0.79).
• On real-world data, 3D-LaneNet+ improves overall MAP to 0.89 and AP50 to 0.95, with AP90 reaching 0.60 using the full model; ablated variants perform worse.
• Removing the global embedding or using a naive clustering reduces MAP and AP90, demonstrating the embedding’s role in accurate lane recovery and longer detections.
• Anchor-free segmentation enables detecting complex topologies such as splits, merges, short lanes, and perpendicular lanes that anchor-based methods miss.
• The method shows strong generalization to new cameras and scenes, including urban junctions, without explicit domain adaptation.
• The real-world dataset results indicate better lateral accuracy (approx. 14–15 cm range) compared with the baseline.