Triple-duty driving perception in one forward pass

What it does

YOLOP is a multi-task perception network for autonomous driving that jointly handles traffic object detection, drivable area segmentation, and lane detection. It uses a shared encoder with separate task heads, training all three end-to-end without alternating optimization. The authors claim it is the first work to achieve real-time performance on embedded devices while maintaining state-of-the-art accuracy on the BDD100K dataset.

The interesting bit

The ablation studies suggest grid-based detection meshes better with semantic segmentation than region-based alternatives. Swapping in an R-CNN-style pipeline cratered both speed (103.3 ms/frame vs. 24.4 ms/frame) and accuracy across every task. Notably, joint training barely hurt single-task performance—detection recall actually improved from 88.2 to 89.2 in the multi-task setup.

Key highlights

• Triple-task network: one forward pass for detection, drivable-area segmentation, and lane detection.
• Reported BDD100K scores: 76.5% mAP50 for detection, 91.5% mIOU for drivable area, and 70.5% mIOU for lanes.
• End-to-end multi-task inference at 24.4 ms/frame (~41 fps), compared with ~15 ms/frame for single-task branches.
• Includes deployment toolkits and demo support for images, video, and camera streams.
• Explicit ablations compare end-to-end, step-by-step, and single-task training strategies.

Caveats

• Several repo components—light.py, postprocess.py, and campus-scene datasets—are explicitly tagged “unrelated to paper,” so the codebase mixes published work with experimental extras.
• Lane detection visualizations rely on post-hoc quadratic fitting, which is not produced by the network itself.
• The README does not specify which hardware was used for the reported speed benchmarks, making the embedded-device performance claim hard to verify from the source alone.

Verdict

A useful research reference if you want to collapse a three-model autonomous-driving perception stack into one efficient network. Skip it if you only need one task, since the single-task ablations show dedicated branches are roughly 1.5× faster.