Day 46 of 133

Object detection (R-CNN family, YOLO, DETR)

Two-stage vs one-stage; how DETR removes anchors and NMS.

DSA · NeetCode Backtracking

SubsetsDSA · Backtracking
Interview questions to prep
1. Compare backtracking vs iterative bitmask vs recursive include/exclude.
2. How does this change for subsets-with-duplicates (LC 90)?
Combination SumDSA · Backtracking
Interview questions to prep
1. How do you avoid duplicate combinations without a set — what's the index-passing trick?
2. Compare with combination-sum-ii where each candidate can only be used once.

R-CNN family, YOLO, DETRDeep LearningLilian Weng
Interview questions to prep
1. Compare two-stage (Faster R-CNN) vs one-stage (YOLO) detectors.
2. How does DETR remove the need for NMS and anchor boxes?
Semantic vs instance segmentation; U-Net & Mask R-CNNDeep LearningU-Net
Interview questions to prep
1. How does U-Net's skip connection design help with segmentation?
2. Compare semantic vs instance vs panoptic segmentation — when do you reach for each?
mAP, IoU, panoptic qualityDeep LearningCOCO
Interview questions to prep
1. Define mAP at IoU=0.5:0.95 — what does each piece mean?
2. Why does NMS need careful tuning, and what does it break when set wrong?
3. How would you implement mAP from predicted boxes, confidences, and class labels?
YOLO anchors, losses, and limitationsDeep LearningYOLO
Interview questions to prep
1. What are anchor boxes in YOLO, and why did anchor-free detectors become popular?
2. What terms usually appear in a YOLO loss, and how do localization, objectness, and class losses interact?
3. Where does YOLO struggle compared with two-stage detectors or DETR?

References & further reading