I have been looking into self-supervised methods in computer vision. For example Preprint Digging Into Self-Supervised Monocular Depth Estimation
which looks at three consecutive frames of video, tasks a network with predicting the third frame, and uses the original as the ground truth / supervisory signal.This type of pretext task for self supervision is a cross between context-based and semantic label based pretext tasks Preprint Self-supervised Visual Feature Learning with Deep Neural Net...
Lane line detection in many ways is approached as a subset of the semantic segmentation task.
I am wondering if there is any way to come up with a pretext task that is specific to lane line detection?
I have seen where self supervision is used in the lane fitting task Chapter Self-supervised Homography Prediction CNN for Accurate Lane ...
But this is used after the lane segments have been identified.
I think that the semantic segmentation approach in this case is better and can be used in real-time in combination with video stream data.
- Badges
- Science topic