Question 1 of 25

Object detection output differs from image classification because detection returns:

Question 2 of 25

Intersection over Union (IoU) for two boxes measures:

Question 3 of 25

Non-maximum suppression (NMS) in object detection:

Question 4 of 25

One-stage detectors (e.g. YOLO) differ from two-stage (Faster R-CNN) mainly by:

Question 5 of 25

mAP (mean Average Precision) in detection:

Question 6 of 25

Feature Pyramid Network (FPN) helps detection by:

Question 7 of 25

Anchor boxes in Faster R-CNN are:

Question 8 of 25

RetinaNet's focal loss addresses:

Question 9 of 25

DETR-style detectors assign predictions to ground truth using:

Question 10 of 25

A typical positive anchor assignment during training uses IoU threshold:

Question 11 of 25

NMS fails in crowded scenes when:

Question 12 of 25

Doubling detector input resolution from 640 to 1280 often:

Question 13 of 25

INT8 quantization for edge detectors:

Question 14 of 25

AP@0.5 vs AP@0.75 differs in that AP@0.75:

Question 15 of 25

Two-stage detectors often have lower false positives on cluttered backgrounds because:

Question 16 of 25

A bounding box stored as (x_center, y_center, width, height) normalized to [0,1] is common because:

Question 17 of 25

Small objects are harder to detect primarily because:

Question 18 of 25

Exporting a detector to ONNX enables:

Question 19 of 25

Lowering the score threshold at inference typically:

Question 20 of 25

When comparing two detectors, reporting only AP@0.5 might:

Question 21 of 25

In torchvision Faster R-CNN, label 0 in `targets["labels"]` means:

Question 22 of 25

When calling `model(images, targets)` in training mode, Faster R-CNN returns:

Question 23 of 25

RoI Align improves on RoI Pool because it:

Question 24 of 25

An empty `boxes` tensor with shape (0, 4) in a target dict is:

Question 25 of 25

Placing NMS on CPU while the model runs on GPU often happens because: