Deep Segmentation MCQ
Segmentation overview, semantic pixel-wise labeling, and instance segmentation with masks.
Segmentation Overview MCQ
Image segmentation overview
Segmentation assigns pixels to coherent regions or objects. Classical methods use intensity/color continuity, edges, or optimize an energy balancing data fidelity and smoothness.
Beyond thresholding
Global thresholds fail under complex scenes—region growing, mean-shift, watershed, and graph cuts add spatial coupling.
Methods
Region growing
Start seeds, merge pixels similar to region statistics—sensitive to seeds and thresholds.
Watershed
Treat gradient magnitude as topography; markers reduce oversegmentation.
Graph cuts
Pixels as nodes, pairwise smoothness + unary data terms; min-cut finds a partition.
Deep segmentation
FCN, U-Net, Mask R-CNN move toward semantic/instance tasks with learned features.
Goal
Pixels → regions or object masks with consistent labels
Semantic Segmentation MCQ
Semantic segmentation
Every pixel is classified (sky, road, person, …) without separating object instances. Fully convolutional networks replaced sliding-window classifiers for dense outputs.
U-Net idea
Skip connections fuse fine spatial detail from the encoder with semantic features from the bottleneck—sharp boundaries.
Building blocks
FCN
Replace FC layers with 1×1 convs; upsample feature maps to input resolution.
ASPP / dilation
Multi-scale context without losing resolution too aggressively.
Loss
Per-pixel cross-entropy; class imbalance may need weighting or focal-style ideas.
vs instance
Semantic does not separate two people as different IDs—instance segmentation adds that.
Output
H×W label map (or H×W×C logits)
Instance Segmentation MCQ
Instance segmentation
Each object instance gets its own mask—even if two people share the class “person.” Often built on top of strong object detectors with a parallel mask branch.
Mask R-CNN
Faster R-CNN backbone + FPN + class/box head + small FCN per ROI to predict a fixed-resolution binary mask.
Details
ROIAlign
Bilinear sampling without harsh quantization—fixes misalignment vs ROI pooling for masks.
AP_mask
Average precision over mask IoU thresholds—COCO-style reporting.
vs semantic
Semantic merges same-class instances; instance keeps them separate.
Alternatives
One-stage instance (YOLACT), query-based (Mask2Former)—tradeoffs in speed vs accuracy.
Output
N instances: class, box, binary mask per instance