1. Detection 모델의 종류
- One-stage detection
- faster & simpler, but lower accuracy
- Two-stage detection:
- slower, but higher accuracy
- First stage: generates a sparse set of candidate object detection
- Two-stage cascade to remove easy negatives
- Second stage: classifies each candidate location as one of the foreground classes or as background using CNN
- Sampling heuristics: foreground-background ratio (1:3), online hard example mining
Must maintain balance between foreground and background samples for dense boxes