LaneDetection_DETR

This is an implementation of lane detection using DETR. Thanks to DETR, End-to-End aproach towards lane detection is achieved. This implementation is based on below contributions.

• original DETR
• transfer learning example of DETR

Here are some examples of inferences.

Requirement

training data

• download dataset of TuSimple lane_detection
• store dataset as follow

.
└── /trainset
    ├── label_data_0313.json
    ├── label_data_0531.json
    ├── label_data_0601.json
    └─ /clips
         ├─ /0313-1
         │    ├─ /60
         │    ....
         ....

environment

This implementation is checked below environment.

• python 3.6.11
• pytorch 1.5.0
• torchvision 0.6.0

Architecture

This is an overview of the network.

Backbone and encoder/decoder is same as original DETR.

Prediction heads are replaced to adopt to lane detection. LaneDetection_DETR outputs lane candidates as belows.

• class label : Same as the original (number of classes + 1) dimension output. This time, it is a two-dimensional output because it was judged only whether it was a lane marking or a background.
• Top and bottom of the lane marking : The y coordinate that shows the target lane marking in the image. Two-dimensional output at the top and bottom.
• X-coordinate sequence of lane markings : The lane markings are regarded as "point sequences separated at equal intervals", and an array containing the x-coordinates of each point is output. An invalid value is set for the height without a lane marking.

Lane candidates are converted to general bounding box in order to judge bipartite matchings between candidates and ground truth.

Brief explanation is written here.