Computer Vision (Q251-290)

Short Answer

Object detection: Finds objects and draws bounding boxes (YOLO, Faster R-CNN). Semantic segmentation: Classifies every pixel (U-Net, DeepLab). Instance segmentation combines both (Mask R-CNN).

Key Points

• Detection: "There's a cat at coordinates (x,y,w,h)"
• Semantic Seg: "These pixels are cat, these are dog"
• Instance Seg: "These pixels are cat #1, those are cat #2"
• Panoptic Seg: Semantic + Instance combined

Short Answer

YOLO (You Only Look Once) divides the image into a grid, predicts bounding boxes and class probabilities in a single forward pass. Unlike two-stage detectors (R-CNN), it's a single-shot detector — making it real-time capable.

Key Points

• Single forward pass: Grid cells predict boxes + confidence + classes
• Non-Maximum Suppression (NMS) removes duplicate detections
• YOLOv8/v9: State-of-the-art, anchor-free, better small object detection
• Trade-off: Slightly less accurate than two-stage but much faster

Short Answer

ViT splits an image into patches, treats each as a token, and applies transformer self-attention. Advantages: global receptive field from layer 1, scales better with data. Disadvantage: needs more data than CNNs (no inductive bias for locality).