Evaluation & Benchmarks MCQ
Detection and segmentation metrics (IoU, mAP), plus ImageNet and COCO benchmark datasets.
CV Evaluation Metrics MCQ
Metrics that match the task
Classification uses accuracy, precision, recall, F1, and ROC-AUC. Object detection pairs predictions to ground truth with IoU thresholds, then averages precision across recall (AP) and classes (mAP). Segmentation reports mean IoU across classes; ranking tasks need NDCG. Always align metric with business cost (false positives vs misses).
Why mAP
Summarizes the quality of ranked detections across IoU cutoffs and recall—better than single-threshold accuracy.
Key ideas
IoU
Intersection over union for boxes or masks.
Precision / Recall
TP/(TP+FP) vs TP/(TP+FN).
AP
Area under precision-recall curve after sorting by score.
mAP
Mean AP over classes (and sometimes IoU thresholds).
Detection match
NMS → sort by score → greedy match to GT with IoU ≥ threshold
ImageNet MCQ
The ImageNet benchmark
ImageNet organizes images under WordNet synsets. The ILSVRC classification track popularized 1000-way supervised training; winning CNNs (AlexNet onward) transferred features to detection, segmentation, and beyond. Download policies and labeling noise are practical considerations when using subsets.
Synset
A synset groups synonymous nouns; each class corresponds to one synset's images.
Key ideas
Hierarchy
WordNet structure relates classes (not flat synonyms only in head).
Scale
Large enough to require strong generalization from CNNs.
ILSVRC
Annual competition that tracked progress on cls/det/seg tasks.
Transfer
Pretrained ImageNet weights initialize many downstream models.
Classification track
1.28M train images, 50k val, 1000 classes (ILSVRC2012 subset)
COCO Dataset MCQ
MS COCO in a nutshell
COCO provides instance segmentation masks, bounding boxes for 80 thing categories, person keypoints, captions, and panoptic labels. Evaluation uses COCO-style mAP with IoU thresholds and area ranges. The COCO API helps load annotations and compute metrics consistently.
Instance vs semantic
COCO instance segmentation separates individual objects with masks; semantic merges same-class regions without object IDs.
Key ideas
Detection
BBox + class with AP@[.5:.95].
Instance seg
Mask AP with mask IoU matching.
Captions
Image → multiple reference sentences; BLEU/CIDEr metrics.
Keypoints
Person skeleton annotations evaluated with OKS-AP.
JSON annotations
images + annotations with bbox, segmentation polygons/RLE, category ids