# Meeting 1 Remarks - Look at slides Lecture 3 again (Most important starts from slide 82) - Make scatter plots of feature pairs, color points by label - For sequences where e.g. hand was undetected, was hand behind head or behind the other hand ? If so, how to deal with occlusion - Undected keypoints should be handled differently per body part based on literature - Look at the sequence(s) with only 1 frame, the one with 117 frames? - Look at sequences more individually overall - Focus on baseline first -> **errors that baseline makes** - Baseline should achieve 0.68 test accuracy - Extensions of baseline should always be better (with new features, undetected keypoint interpolation, ...) - Look at errors thoroughly - Most important features of baseline ? - Look at correlation between features (-> scatter plots) - Only from observations of error results you can try new things (i.e. don't try to transform features at random like we did before), in other words you need to have a reason and grounds before implementing things - Add new features to baseline -> should always perform better - Confusion matrices not enough - Examine samples where model is confident but wrong - Examine samples for which model has very low confidence - Focus on confused classes - Most frequently confused class pairs - Which features are selected by current linear model: mouth, hands, body ?