A region proposal approach makes object detection manageable by first generating candidate image regions, then predicting objects within those regions. It supports a variable number of predictions and pairs naturally with easier-to-label bounding boxes. section 8.0
The practical tradeoff is:
- Bounding-box labels reduce annotation effort because a labeler only clicks and drags a box around each object. section 8.0
- Region proposal detection is widely used and available as a built-in PyTorch component, while implementation-specific details can be deferred. section 8.0
Bounding-box detection
Instead of producing one prediction for every pixel, a detector proposes regions that may contain objects and produces a variable number of bounding-box predictions. section 8.0
