March 2021
tl;dr: Extend yolo to perform single-stage instance segmentation.
Insta-yolo adopts a fixed length contour representation, and uses a
Work by Valeo Egypt. Speed is very fast but quality is subpar. Looks like a run-of-the-mill paper.
- Represent masks by a fixed number of contour points (polygons) in Cartesian, and predict the polygons of each object instance through the center of the object.
- GT generation with a deterministic algorithm (dominant points detection).
- Loss
- Regression loss wrt the GT generated with deterministic algo
- IoU Loss to compensate for the fact that no unique representation for the object mask using fixed number of vertices.
- This can also work for orientated bbox prediction.
- Log Cosh loss: a differentiable alternative to Huber loss (smooth L1 loss).