Shir Bar - Lecture 6 - Evaluation Metrics in Computer Vision for Ecology

Metrics are mathematical statements of our goals. Training our model will maximize this goal! Afterwards, we use metrics to compare different approaches to our problem. The lecturer, Shir Bar from Tel Aviv University, will introduce common metrics in Classification, Segmentation, and Detection. She will also detangle the Confusion Matrix, Precision-Recall Curve, and why Accuracy is so misleading in imbalanced data settings. ⛆ Contents ⛆ 🐡 (00:03) Think-Pair-Share Activity 🐡 (07:02) Introduction 🐡 (09:52) Review of Computer Vision Tasks 🐡 (11:35) Building Blocks: Discrete Hamming distance, Continuous MSE, Spatial IoU 🐡 (14:58) Binary Classification: Intro 🐡 (17:13) Binary Classification: Confusion Matrix, Precision, and Recall 🐡 (19:00) Binary Classification: Selecting the decision threshold and P-R Curve 🐡 (27:00) Binary Classification: F-Score 🐡 (27:28) Multi-Class: Top-k accuracy 🐡 (30:15) Class Imbalance: Macro-averaging, class-balanced metric 🐡 (40:08) Semantic segmentation: IoU 🐡 (44:32) Object Detection: P-R Curve, AP, mAP 🐡 (56:57) Food for Thought