Generating bird’s-eye view from multi-image scene using self-supervised learning

  • Utilized a pre-trained denoising auto-encoder as a feature extractor for the task of road map construction.
  • Combined the feature embeddings of the multi-view scene captured from moving vehicles and used a U-Net to generate a top-down view of the binary road map.
  • Achieved a threat score of 0.70 on the hold-out test set (with the best score in the intra-class competition being ~0.81).
Mrinal Jain
Mrinal Jain
Student

Trying to make silicon do what carbon can.