Segmentation and Pose
Semantic, instance, and panoptic segmentation, their CE/Dice/IoU losses and metrics, and keypoint pose estimation.
Key Terms
Semantic SegmentationInstance / PanopticCross-Entropy / Dice LossmIoU / PQKeypointsHeatmap Regression
Python Tools
PyTorchOpenCVNumPyMatplotlib
Textbook Reference
Szeliski Ch.6
Topics Covered
- 1Task definitions: semantic, instance, and panoptic segmentation
- 2Architectures: encoder-decoder, FCN, U-Net, and Mask R-CNN
- 3Losses: pixel cross-entropy, Dice, and boundary-aware terms
- 4Metrics: pixel accuracy, mean IoU, and panoptic quality
- 5Human pose estimation: heatmap regression, top-down versus bottom-up, and OKS
Learning Outcomes
- Distinguish semantic, instance, and panoptic segmentation and their metrics
- Choose segmentation losses and explain the IoU and Dice trade-offs
- Describe heatmap-based keypoint estimation and its evaluation