zkyseu/o2sformer

How zkyseu/o2sformer works

O2SFormer is a PyTorch lane-detection system for CULane. The user workflow is to train from configured dataset splits, evaluate predictions with the CULane metric, or load a checkpoint and run image inference for visualization. Its model implementation includes detector variants, lane-prediction heads, transformer components, assignment strategies, and losses; the README also documents a HybridEncoder variant. Several internal model connections are supported by the source index, but the complete detector-to-head wiring is not shown in the excerpts, so that relationship is left unasserted.

Main entry points: culane_dataset.py CULane dataset · transformer.py Transformer · detection_head.py Detection head · detection_headv2.py Detection head v2

Components 24

Training and inference

Training API dnlane/apis/train.py
Inference API dnlane/apis/inference.py
Model configs configs
Evaluation hook dnlane/hook/eval_hook.py
Lane predictions

CULane data and evaluation

CULane dataset dnlane/datasets/culane_dataset.py
Lane formatting dnlane/datasets/pipelines/lane_formating.py
Image transforms dnlane/datasets/pipelines/transform.py
CULane metric dnlane/datasets/culane_metric.py

Lane detection model

Detector variants dnlane/models/detector.py
Detector v2 dnlane/models/detectorv2.py
Detection head dnlane/models/detection_head.py
Detection head v2 dnlane/models/detection_headv2.py

Model components

Lane representation dnlane/models/lane.py
Transformer dnlane/models/transformer.py
Transformer utilities dnlane/models/transformer_utils.py
Hybrid encoder dnlane/models/hybrid_encoder.py
Assignment strategies dnlane/models/assigner.py
One-to-many assigner dnlane/models/o2m_assigner.py
Match costs dnlane/models/match_cost.py
Lane IoU loss dnlane/models/losses/lane_iou.py
Focal loss dnlane/models/losses/focal_loss.py
General utilities dnlane/models/utils/general_utils.py

Other components

User

Connections 20

User → Training API: starts training README.md
User → Inference API: requests inference
Model configs → Training API: configures training README.md
Inference API → Detector variants: builds detector dnlane/apis/inference.py
Inference API → Lane predictions: returns predictions dnlane/apis/inference.py
CULane dataset → Lane formatting: runs pipeline dnlane/datasets/culane_dataset.py
Lane formatting → Image transforms: uses transforms dnlane/datasets/pipelines/lane_formating.py
CULane dataset → CULane metric: evaluates predictions dnlane/datasets/culane_dataset.py
Evaluation hook → CULane dataset: tests dataloader dnlane/hook/eval_hook.py
Detector variants → Transformer: uses transformer dnlane/models/detector.py
Detector v2 → Transformer: uses transformer dnlane/models/detectorv2.py
Detection head → Transformer: uses transformer dnlane/models/detection_head.py
Detection head → Transformer utilities: uses utilities dnlane/models/detection_head.py
Detection head → Lane representation: uses lane type dnlane/models/detection_head.py
Detection head v2 → Transformer utilities: uses utilities dnlane/models/detection_headv2.py
Detection head v2 → Lane representation: uses lane type dnlane/models/detection_headv2.py
Detection head v2 → Lane IoU loss: uses IoU loss dnlane/models/detection_headv2.py
Hybrid encoder → Transformer: uses transformer dnlane/models/hybrid_encoder.py
One-to-many assigner → Lane IoU loss: uses IoU dnlane/models/o2m_assigner.py
CULane dataset → General utilities: uses path helper dnlane/datasets/culane_dataset.py

Drawn by GitDiagram on November 12, 2026 from the repository's file tree, README and source files. Read as Markdown ·