Back to Subject

Unit 4: Object Detection and Localization - Practice Quiz

CSE471 — Deep Learning For Computer Vision 60 Questions
0 Correct 0 Wrong 60 Left
0/60

1 What is the main goal of object detection in an image?

Object detection principles Easy
A. Remove noise from the image
B. Locate and classify objects
C. Increase the image resolution
D. Classify the entire image

2 How does object detection differ from image classification?

Object detection principles Easy
A. It predicts labels and locations
B. It converts images into grayscale before assigning a single category
C. It only changes image colors
D. It only assigns one image label

3 Which output is typically produced by an object detection algorithm?

Object detection algorithms Easy
A. A class label and bounding box
B. A grayscale intensity value
C. A complete textual description containing every visible detail in the image
D. A compressed image file

4 Which set of values can represent a rectangular bounding box?

Bounding box representation Easy
A. Width, depth, and brightness
B. Class, height, and angle
C. Center coordinates, width, and height
D. Center, width, and color

5 In the corner-based representation , what does usually indicate?

Bounding box representation Easy
A. The average coordinate of all pixels inside the box
B. The box center
C. The bottom-right corner
D. The top-left corner

6 What does Intersection over Union (IoU) measure?

Intersection over union Easy
A. The overlap between two boxes
B. The number of object classes
C. The brightness of an image
D. The training time of a detector

7 Which formula defines Intersection over Union for two bounding boxes?

Intersection over union Easy
A.
B.
C.
D.

8 What is the IoU of two identical bounding boxes?

Intersection over union Easy
A.
B.
C.
D.

9 Why is non-maximum suppression used in object detection?

Non-maximum suppression Easy
A. To enlarge small objects
B. To remove duplicate detections
C. To create additional object classes
D. To replace every predicted box with a detailed pixel-level segmentation mask

10 Which bounding box is usually kept first during non-maximum suppression?

Non-maximum suppression Easy
A. The box with the smallest area
B. The box nearest the image corner
C. The box with the highest confidence
D. The box with the largest width

11 What does a sliding-window detector do?

Classical detection pipelines Easy
A. Generates a detailed natural-language caption for each pixel in the image
B. Rotates the image once
C. Labels only the full image
D. Examines many image regions

12 Why are image pyramids used in classical object detection?

Classical detection pipelines Easy
A. To detect objects at different scales
B. To assign one identical bounding box to every object in an image
C. To convert objects into triangles
D. To reduce the number of classes

13 What is used by the original R-CNN to identify candidate object regions?

RCNN Easy
A. Randomly generated labels
B. Image captions
C. Audio samples
D. Region proposals

14 What does the abbreviation R-CNN stand for?

RCNN Easy
A. Region-based Color Normalization Network
B. Recursive Convolutional Number Network
C. Region-based Convolutional Neural Network
D. Regional Classification Neural Network

15 What characterizes a single-stage object detector?

Single-stage detectors Easy
A. It processes only grayscale images
B. It requires manual box drawing
C. It predicts detections directly
D. It creates region proposals in one network and classifies them using a separate multi-step pipeline

16 What does YOLO stand for in object detection?

YOLO variants Easy
A. Yielding Optimized Layer Operations
B. Your Object Localization Output
C. You Only Label Objects
D. You Only Look Once

17 Which property is commonly associated with YOLO variants?

YOLO variants Easy
A. Fast object detection
B. Audio signal generation
C. Manual feature labeling
D. Separate processing of every proposed region through several independent classifiers

18 What does SSD stand for in computer vision?

SSD Easy
A. Single Shot Detector
B. Standard Scale Detection
C. Spatial Search Decoder
D. Sequential Sampling Device

19 What problem is focal loss designed to reduce in RetinaNet?

RetinaNet with focal loss Easy
A. The need to store every training image at several different resolutions
B. Box visualization
C. Image rotation
D. Class imbalance

20 What does mean average precision (mAP) summarize in object detection?

Mean average precision Easy
A. Maximum number of boxes
B. Total training duration
C. Detection performance across classes
D. Average image brightness

21 An image contains three cars and two pedestrians. What must an object detector produce that an image classifier typically does not?

Object detection principles Medium
A. A feature vector representing the complete image
B. A class label and location for each object instance
C. A single class label for the complete image
D. A pixel-level class label for every image pixel

22 A bounding box is represented as . What are its width and height?

Bounding box representation Medium
A. Width , height
B. Width , height
C. Width , height
D. Width , height

23 Two predicted boxes have an intersection area of pixels. Their individual areas are and pixels. What is their IoU?

Intersection over union Medium
A.
B.
C.
D.

24 A predicted box is completely inside a ground-truth box. Their areas are and pixels, respectively. What is their IoU?

Intersection over union Medium
A.
B.
C.
D.

25 For one class, three boxes have confidence scores , , and . The IoU of the first box with the other two is and , respectively. With an NMS threshold of , which boxes are retained?

Non-maximum suppression Medium
A. The boxes scored and
B. All three candidate boxes
C. The boxes scored and
D. Only the box scored

26 What is the most likely effect of increasing the NMS IoU threshold from to while keeping all confidence scores unchanged?

Non-maximum suppression Medium
A. More highly overlapping boxes will be retained
B. Bounding boxes will become geometrically smaller
C. More low-confidence boxes will be removed
D. Fewer object classes will be evaluated

27 In a sliding-window detector, reducing the stride from pixels to pixels generally has which effect?

Classical detection pipelines Medium
A. It evaluates fewer windows and decreases computation
B. It removes the need for an image pyramid
C. It converts the detector into a segmentation model
D. It evaluates more windows and increases computation

28 Why is an image pyramid commonly used with a fixed-size sliding-window classifier?

Classical detection pipelines Medium
A. To eliminate overlapping candidate windows
B. To compute exact object segmentation masks
C. To assign multiple labels to one window
D. To detect objects appearing at different scales

29 Which operation is a major source of slow inference in the original R-CNN architecture?

RCNN Medium
A. Computing all convolutional features only once
B. Predicting boxes directly from a dense feature map
C. Sharing one classifier across all image locations
D. Running the CNN separately on each region proposal

30 How does Fast R-CNN reduce repeated computation compared with the original R-CNN?

RCNN Medium
A. It computes a shared feature map before processing regions
B. It performs classification without bounding-box regression
C. It applies a separate backbone to every proposed region
D. It replaces region proposals with an image pyramid

31 What is the main contribution of Faster R-CNN over Fast R-CNN?

Object detection algorithms Medium
A. It learns region proposals using a Region Proposal Network
B. It removes bounding-box regression from the detection head
C. It replaces convolutional features with handcrafted descriptors
D. It classifies only one object in each input image

32 Why are single-stage detectors often faster than two-stage detectors?

Single-stage detectors Medium
A. They use ground-truth boxes as proposals during inference
B. They evaluate every region using an independent backbone
C. They avoid convolutional feature extraction during inference
D. They predict classes and boxes without a separate proposal stage

33 A YOLO-style detector divides an image into a grid. In the standard assignment principle, which grid cell is responsible for predicting an object?

YOLO variants Medium
A. The cell having the largest image area
B. The cell containing the object's top-left corner
C. The cell containing the object's center
D. The cell having the lowest class confidence

34 A detector performs well on large objects but frequently misses small objects. Which feature found in newer YOLO variants most directly addresses this issue?

YOLO variants Medium
A. Classification from only the deepest feature map
B. Use of one bounding box for the full image
C. Removal of all high-resolution backbone features
D. Predictions from multiple feature-map scales

35 Why does SSD make predictions from several feature maps with different spatial resolutions?

SSD Medium
A. To convert detection into image classification
B. To detect objects across a range of sizes
C. To guarantee one prediction for each object
D. To avoid using default boxes during training

36 During SSD training, how is a ground-truth box typically matched to default boxes?

SSD Medium
A. By selecting boxes with the lowest class confidence
B. By selecting default boxes with sufficient IoU overlap
C. By selecting boxes from the deepest feature map only
D. By selecting boxes with identical pixel coordinates only

37 In a dense detector, most anchors are easy background examples. How does focal loss mainly address this imbalance?

RetinaNet with focal loss Medium
A. It assigns equal loss to all positive and negative anchors
B. It removes every background anchor before training
C. It increases the loss from easy negative examples
D. It down-weights easy, confidently classified examples

38 For focal loss with focusing parameter , what happens to the modulating factor for a correctly classified example as its predicted probability approaches ?

RetinaNet with focal loss Medium
A. It becomes negative, reversing the loss gradient
B. It approaches , preserving the example's full contribution
C. It grows without bound, dominating the total loss
D. It approaches , reducing the example's contribution

39 A detector produces one correct high-confidence prediction for every ground-truth object, followed only by several low-confidence false positives. Why can its average precision still be high?

Mean average precision Medium
A. Average precision measures localization error only
B. False positives are ignored at every confidence threshold
C. Correct detections appear early in the confidence ranking
D. Low-confidence predictions automatically become true negatives

40 The same detector is evaluated first at IoU threshold and then at . What outcome is most likely when the threshold is increased?

IoU thresholds Medium
A. Some loosely localized detections become false positives
B. More detections qualify as correct localizations
C. Non-maximum suppression becomes unnecessary
D. Class confidence scores automatically increase

41 An image contains one non-ignored ground-truth object. A detector outputs three correctly classified boxes with IoUs , , and in descending confidence order. At an IoU threshold of , how are these detections labeled during standard one-to-one evaluation?

Object detection principles Hard
A. The first is a true positive, the second is ignored, and the third is a false positive.
B. The first is a true positive, and the other two are false positives.
C. All three are true positives because each predicts the correct class.
D. The first two are true positives, and the third is a false positive.

42 An anchor has center , width , and height . A detector uses , , , and . For predicted offsets , what decoded corner box is obtained?

Bounding box representation Hard
A.
B.
C.
D.

43 Using continuous-coordinate box areas, let and , where each tuple is . What is ?

Intersection over union Hard
A.
B.
C.
D.

44 Three same-class detections have scores , , and . Their pairwise IoUs are , , and . With greedy NMS suppressing IoU values greater than , which boxes remain?

Non-maximum suppression Hard
A. and
B. and
C. and
D. only

45 A HOG-SVM pedestrian detector was trained only on fixed-size positive windows. At test time, pedestrians may appear at many scales. Which pipeline most directly preserves the classifier's expected window representation while enabling multiscale detection?

Classical detection pipelines Hard
A. Use one global HOG descriptor and infer every pedestrian box from its SVM margin.
B. Resize every candidate window independently to the full image dimensions, classify all pixels jointly, and average overlapping scores before thresholding.
C. Scan variable-sized windows on one image scale without resizing their descriptors.
D. Scan a fixed window over an image pyramid, map detections back, then apply NMS.

46 Which change specifically distinguishes Faster R-CNN from Fast R-CNN while retaining a two-stage detection structure?

RCNN Hard
A. It runs the convolutional backbone separately for every proposed region.
B. It removes region proposals and predicts all boxes directly from a dense output grid.
C. It replaces external region proposals with a learned region proposal network sharing convolutional features.
D. It replaces ROI feature extraction with class-agnostic NMS during training.

47 Why can a dense single-stage detector perform poorly when trained with ordinary cross-entropy over all anchors, even if most background anchors are individually easy?

Single-stage detectors Hard
A. Dense prediction prevents the backbone from sharing features across spatial positions.
B. Cross-entropy produces zero gradients for all correctly classified background anchors.
C. Background anchors are always assigned larger regression targets than positive anchors.
D. The aggregate gradient from numerous easy negatives can dominate the sparse positive signal.

48 In a YOLO-style detector that assigns each ground-truth object to the grid cell containing its center, an object's center lies just inside one cell while most of its area extends into neighboring cells. Which statement is correct?

YOLO variants Hard
A. The object is discarded because its box is not fully contained in one grid cell.
B. Every intersected cell must independently predict the complete object as a positive.
C. The center-containing cell can regress a box extending well beyond its own boundaries.
D. The box must be clipped to the responsible cell before computing localization loss.

49 In standard SSD matching, a ground-truth box has IoUs , , and with its three best default boxes, while the normal positive threshold is . What prevents this ground truth from necessarily having zero positive matches?

SSD Hard
A. Its highest-IoU default box is force-matched to it before threshold-based matching.
B. The positive threshold is automatically reduced to the mean of the three IoUs.
C. All default boxes with IoU above become positives for that ground truth.
D. The ground truth is converted into a default box and appended to the prediction tensor.

50 For positive examples, focal loss is . With and the same , approximately what is ?

RetinaNet with focal loss Hard
A.
B.
C.
D.

51 A class has two ground-truth objects. In descending confidence order, three detections are labeled TP, FP, TP. Using all-point interpolated precision and the area under the precision-recall curve, what is the AP?

Mean average precision Hard
A.
B.
C.
D.

52 Two detectors have equal , but detector has substantially higher COCO-style AP averaged over IoU thresholds from to . What is the strongest justified conclusion?

IoU thresholds Hard
A. Detector produces fewer raw candidate boxes before confidence filtering and NMS at every tested image resolution.
B. Detector must have higher recall when detections are evaluated only at IoU .
C. Detector necessarily detects more object classes than the other detector.
D. Detector generally localizes objects more precisely at stricter overlap requirements.

53 A proposal generator has recall at IoU with the ground truth. Its downstream classifier and box head are otherwise perfect but may only process those proposals. What is the maximum achievable detection recall at IoU ?

Object detection algorithms Hard
A.
B.
C.
D.

54 Why are width and height commonly encoded as and relative to an anchor rather than as unrestricted additive pixel offsets?

Bounding box representation Hard
A. The encoding makes the IoU exactly linear in the predicted width and height values.
B. The encoding eliminates the need to normalize center-coordinate regression targets.
C. The encoding guarantees that every decoded box remains inside the image boundary.
D. The encoding models multiplicative scale changes and guarantees positive decoded sizes.

55 Two adjacent objects of the same class genuinely overlap with IoU , and the detector predicts one accurate high-score box for each. Hard class-wise NMS uses threshold . What failure is most likely?

Non-maximum suppression Hard
A. Both detections are relabeled as background before classification.
B. Neither detection is affected because they correspond to different instances.
C. Both detections are merged into a new averaged box, reducing precision.
D. One correct detection is suppressed, reducing recall.

56 For a binary example with label , suppose its predicted probability approaches . Compared with ordinary cross-entropy, what does focal loss with do asymptotically?

RetinaNet with focal loss Hard
A. It converts the classification loss into a bounding-box IoU loss near convergence.
B. It keeps the example's gradient constant so rare classes cannot be forgotten.
C. It increases the loss because diverges as approaches .
D. It suppresses the easy example's contribution by an additional factor .

57 For one class, a detector produces the same set of boxes and confidence values before and after a calibration change, but the change reverses the score order of one true positive and one false positive. Why can AP change even when the final counts of true positives and false positives do not?

Mean average precision Hard
A. AP treats every lower-confidence detection as ignored after the first false positive.
B. AP depends on the ranked precision-recall trajectory, not only on final counts.
C. AP recomputes box coordinates from confidence scores before calculating IoU.
D. AP changes the number of ground-truth objects according to the score ordering.

58 A predicted square has the correct center and side length , while the ground-truth square has side length . Both are axis-aligned. What is their IoU, and at which listed evaluation thresholds is the prediction a true positive?

IoU thresholds Hard
A. IoU ; it is a true positive at all thresholds up to .
B. IoU ; it is a true positive only at threshold .
C. IoU ; it is a true positive at thresholds and .
D. IoU ; it is a true positive at thresholds and .

59 A YOLO head predicts an objectness value intended to represent the probability that an object is present and a conditional class distribution. Under this factorization, which score is appropriate for ranking detections of class before NMS?

YOLO variants Hard
A.
B.
C.
D.

60 An SSD model uses default boxes with many aspect ratios but only coarse, low-resolution feature maps. Which objects are most likely to remain difficult despite the aspect-ratio diversity?

SSD Hard
A. Large objects whose aspect ratios exactly match multiple default boxes.
B. Very small objects whose centers and features are poorly resolved spatially.
C. Objects from frequent classes whose hard-negative examples are selected during loss computation.
D. Medium objects that overlap several adjacent default boxes above the matching threshold.
Previous MCQ
Unit 3
Next MCQ
Unit 5