Deep Learning in Computer Vision Practice Assignment

Fill-in-the-blank practice flashcards covering RANSAC, homography, CNN fundamentals, object detection models like YOLO and Mask R-CNN, and tracking-by-detection concepts.

The main purpose of RANSAC in feature matching is to reject geometrically inconsistent __________.

matches

An inlier in RANSAC is a point or match that agrees with the estimated model within a __________.

threshold

For homography estimation using RANSAC, the minimum number of point correspondences sampled in one trial is __________.

$$4$$

A homography is mainly used to map points between two image __________.

planes

A homography is represented by a $3 \times 3$ matrix with __________ degrees of freedom.

$$8$$

The main difference between classical filters and CNN filters is that CNN filters are learned from __________ during training.

data

In a convolutional layer, each trainable kernel produces one __________.

feature map

__________ controls how far the kernel moves at each step during the convolution process.

Stride

__________ is used in CNNs to allow boundary pixels to be processed and to control the spatial size of the output.

Padding

The __________ activation function adds non-linearity by setting all negative values to zero.

ReLU

__________ average pooling converts each complete feature map into one single average value.

Global

During CNN training, __________ is responsible for computing how weights contributed to the total error.

backpropagation

Data augmentation is used to increase training-data variation and improve model __________.

generalization

Dropout helps reduce __________ by randomly setting some activations to zero during training.

overfitting

YOLO is categorized as a __________ detector because it predicts boxes and classes directly in one forward pass.

single-stage

The three common components of the YOLO architecture are the backbone, neck, and __________.

detection head

Non-Maximum Suppression (NMS) is used in object detection to remove overlapping duplicate detections and keep the __________ box.

strongest

Mask R-CNN extends object detection by adding a segmentation __________ for each detected instance.

mask

SAM is a promptable model that focuses on generating segmentation masks from prompts such as __________ or boxes.

points

The metric __________ measures the overlap between a predicted box and the ground truth.

IoU

In a tracking-by-detection pipeline, the detector first predicts __________ in each video frame.

boxes

Unlike a tracker using only box overlap, __________ uses appearance features to compare how objects look to reduce identity switches.

DeepSORT

If a system needs to mark the exact pixels of a damaged region, __________ is more suitable than simple object detection.

segmentation

Two-stage detectors, such as __________, first generate candidate object regions before classifying them.

Faster R-CNN

When evaluating a model on a dataset with a rare class, metrics such as precision, __________, and F1-score are more informative than accuracy.

recall