Why should a detection pipeline name the metric used to save its best checkpoint?
Computer VisionLevel: IntermediateExplicit checkpoint selection metrics
Why should a detection pipeline name the metric used to save its best checkpoint?
Hints
- Which evaluation key does automatic selection actually use?
- Checkpoint selection and final test evaluation have different roles.
Answer
Automatic selection may use the first returned evaluation key, which need not match the intended objective. Select a meaningful validation metric such as bbox or mask mAP and its direction explicitly. Keep validation cadence and evaluation settings fixed. The final test set should evaluate the chosen checkpoint, not repeatedly choose among checkpoints during development.
Source and licence: OpenMMLab — MMDetection FAQ · Apache-2.0 · answer: Quantoros editorial