Computer Vision - MCQ Practice Questions

Pooling (e.g., Max Pooling or Average Pooling) is used in CNNs to reduce the spatial dimensions (height and width) of feature maps. This reduces computation, controls overfitting, and provides a form of translation invariance. Dropout is a regularization technique, Batch Normalization normalizes activations, and Flattening converts a multi-dimensional tensor into a 1D vector before fully connected layers.

IoU (Intersection over Union) is a metric used to evaluate object detection algorithms. It is calculated as the area of overlap (intersection) between the predicted bounding box and the ground truth bounding box divided by the area of their union. A higher IoU indicates a more accurate detection. Option C describes Recall, and Option A describes pixel accuracy used in segmentation.

In U-Net, skip connections concatenate feature maps from the encoder (contracting path) directly to the corresponding decoder (expanding path) layers. This allows the decoder to recover fine-grained spatial details that are lost during downsampling, which is crucial for precise pixel-wise segmentation. They do not replace pooling or affect the learning rate.

The Harris Corner Detector uses the eigenvalues of the second-moment matrix $M$ to classify image regions. If both ${\lambda }_1$ and ${\lambda }_2$ are small, the region is flat. If one is large and one is small (${\lambda }_1\gg {\lambda }_2$ or vice versa), the region is an edge. If both ${\lambda }_1$ and ${\lambda }_2$ are large, the region is a corner. The Harris response is $R=\det (M)-k\cdot (\text{trace}(M){)}^2$, and $R>0$ (large positive) indicates a corner.

Edge detection algorithms like Canny or Sobel compute image gradients, which are highly sensitive to noise. Applying a Gaussian filter (a smoothing/blurring operation) before edge detection reduces high-frequency noise in the image. This prevents noise pixels from being incorrectly identified as edges, resulting in cleaner and more accurate edge maps.

SIFT (Scale-Invariant Feature Transform) works in two main stages: (1) Keypoint detection using the Difference of Gaussians (DoG) approximation of the Laplacian of Gaussian (LoG) across multiple scales, and (2) Descriptor creation using 128-dimensional histograms of gradient orientations in a 16x16 neighborhood around each keypoint. SIFT is invariant to scale, rotation, and partially to illumination. It uses floating-point descriptors, unlike BRIEF or ORB which use binary descriptors. It is computationally expensive, not suited for real-time use.

Semantic segmentation assigns a class label to every pixel in an image (e.g., all cars are labeled 'car'), but it does not differentiate between individual objects of the same class. Instance segmentation, on the other hand, not only classifies each pixel but also distinguishes between separate instances of the same class (e.g., Car 1, Car 2). Panoptic segmentation combines both. Semantic segmentation works on color images, not just binary images, and uses pixel masks, not bounding boxes.

The Hough Transform is a feature extraction technique used to detect parametric shapes — shapes that can be described by a mathematical equation with parameters. The classical Hough Transform detects straight lines (parameterized by slope and intercept, or $\rho$ and $\theta$). It has been extended to detect circles (Circle Hough Transform) and ellipses. It is not limited to rectangles or irregular blobs, and does not handle Bezier curves directly.

In a GAN framework, there are two networks trained adversarially. The Generator takes random noise as input and tries to generate realistic images. The Discriminator takes an image as input and tries to classify it as either 'real' (from the training dataset) or 'fake' (generated by the generator). The generator improves by trying to fool the discriminator, while the discriminator improves by becoming better at distinguishing real from fake images. The encoder role belongs to VAEs (Variational Autoencoders).

Non-Maximum Suppression (NMS) is a post-processing step in object detection pipelines (e.g., YOLO, Faster R-CNN). Object detectors often generate multiple overlapping bounding boxes for the same object. NMS works by selecting the bounding box with the highest confidence score and suppressing (removing) all other boxes that have an IoU greater than a predefined threshold with the selected box. This process is repeated until no overlapping boxes remain, resulting in a single clean detection per object.