The IIT Madras BS Deep Learning for Computer Vision (Deep Learning for Computer Vision) End Term paper sat on 28 Apr 2024, in the January 2024 term: 44 questions for 72 marks in 180 minutes. Every question is below with its answer. Take it as a timed mock test to be marked, or read it through first.
Given word embeddings for two words and weight matrix to generate queries, keys and values as below (in practice, these would be more complex or learned; for simplicity, we provide them here):
Calculate the query, key and value vectors for each word?
Correct answer
What is the role of the Gaussian blur in the Canny edge detection algorithm, and how does the choice of the standard deviation parameter affect the performance of edge detection? Which option do you think best describes the role of the Gaussian blur and its relationship with the standard deviation parameter in the Canny edge detection algorithm?
The Gaussian blur helps to reduce noise in the image and smooth out pixel intensity variations. The standard deviation parameter controls the amount of blurring applied, with higher values resulting in more aggressive smoothing, which may lead to loss of edge detail.
The Gaussian blur enhances edge contrast in the image and amplifies pixel intensity gradients. The standard deviation parameter determines the width of the blur kernel, with larger values resulting in sharper edges and finer details preserved.
The Gaussian blur serves to highlight high-frequency components in the image and accentuate edge boundaries. The standard deviation parameter dictates the level of detail preserved, with smaller values preserving finer details but potentially amplifying noise.
The Gaussian blur functions to dilate edge regions in the image and expand the edge boundaries. The standard deviation parameter controls the extent of dilation, with higher values resulting in broader edge regions but also increasing the risk of false edge detections.
Correct answer
The Gaussian blur helps to reduce noise in the image and smooth out pixel intensity variations. The standard deviation parameter controls the amount of blurring applied, with higher values resulting in more aggressive smoothing, which may lead to loss of edge detail.
Which property of the SIFT descriptor is primarily responsible for its resilience to changes in scale, viewpoint, and illumination?
Scale space representation
Histogram of gradients
Keypoint detection
Orientation assignment
Correct answer
Scale space representation
In the Maximally Stable External Regions (MSER) algorithm, how does the tree structure contribute to the efficiency of region stability analysis?
By enabling quick access to all regions in constant time.
By allowing for parallel computation of stability measures across regions.
By organizing regions hierarchically based on stability, facilitating efficient stability analysis.
By providing a compact representation of pixel intensities within each region, reducing computational overhead.
Correct answer
By organizing regions hierarchically based on stability, facilitating efficient stability analysis.
In which scenarios is the Random Sample Consensus (RANSAC) algorithm more suitable compared to least squares?
When the dataset contains a high percentage of outliers or erroneous data points.
When the dataset consists of noise-free and well-behaved data with minimal outliers.
When the computational resources are limited, and quick solutions are required.
When the dataset is small, and exact solutions are necessary without considering outliers.
Correct answer
When the dataset contains a high percentage of outliers or erroneous data points.
In what way does VLAD (Vector of Locally Aggregated Descriptors) differ from Bag of Words (BoW) in representing visual features in images?
VLAD captures spatial relationships between local descriptors, while BoW does not.
BoW captures spatial relationships between local descriptors, while VLAD does not.
VLAD and BoW both capture spatial relationships between local descriptors.
Neither VLAD nor BoW capture spatial relationships between local descriptors.
Correct answer
VLAD captures spatial relationships between local descriptors, while BoW does not.
How does pyramid matching differ from Speeded-Up Robust Features (SURF) in image matching?
Pyramid matching matches images at multiple resolutions using multi- resolution representations, while SURF focuses on detecting and describing local features invariant to scale and rotation.
SURF relies on gradient-based descriptors for feature matching, whereas pyramid matching utilizes histogram-based representations.
Pyramid matching is computationally more efficient compared to SURF, which requires extensive feature extraction and matching procedures.
SURF is more suitable for matching images with complex patterns and cluttered backgrounds, while pyramid matching excels in scenarios where multi-scale matching is critical, such as image registration across different resolutions.
Correct answer
Pyramid matching matches images at multiple resolutions using multi- resolution representations, while SURF focuses on detecting and describing local features invariant to scale and rotation.
In the context of Convolutional Neural Networks (CNNs), what is the primary challenge associated with backpropagation compared to traditional feedforward neural networks?
The increased number of parameters and layers in CNNs make
backpropagation computationally intensive and prone to overfitting.
The presence of convolutional and pooling layers in CNNs requires the development of specialized backpropagation algorithms for efficient gradient computation.
The non-linear activation functions used in CNNs introduce discontinuities in the error surface, making it difficult to find the global minimum during backpropagation.
The spatial structure of CNNs results in weight sharing and local connectivity, requiring careful consideration of the error propagation process during backpropagation.
Correct answer
The spatial structure of CNNs results in weight sharing and local connectivity, requiring careful consideration of the error propagation process during backpropagation.
What is the primary advantage of Depthwise Separable Convolutions compared to standard Convolutional Neural Network (CNN) convolutions?
Depthwise Separable Convolutions reduce computational complexity and memory footprint by separating the spatial and channel-wise convolutions.
Depthwise Separable Convolutions introduce additional parameters to capture complex spatial relationships within feature maps.
Depthwise Separable Convolutions increase model capacity by performing convolutions across multiple layers of the network.
Depthwise Separable Convolutions enhance feature representation by incorporating multiscale convolutions within the same layer.
Correct answer
Depthwise Separable Convolutions reduce computational complexity and memory footprint by separating the spatial and channel-wise convolutions.
Which of the following strategies can help mitigate the risk of overfitting when fine-tuning pretrained models on a small dataset?
Increasing the depth and width of the neural network architecture to capture more complex patterns in the data.
Applying dropout regularization to randomly deactivate neurons during training to prevent co-adaptation of features.
Decreasing the learning rate during fine-tuning to ensure more stable convergence and smoother optimization.
Increasing the batch size during fine-tuning to improve the stability of gradient estimation and accelerate convergence.
Correct answer
Applying dropout regularization to randomly deactivate neurons during training to prevent co-adaptation of features.
What is the primary advantage of visualizing patches of maximally activating neurons in neural network interpretation?
It helps identify the specific input features that contribute most strongly to the model’s predictions.
It provides insights into the architectural design choices made in the neural network.
It quantifies the uncertainty associated with the model’s predictions.
It measures the computational resources required to activate the neurons.
Correct answer
It helps identify the specific input features that contribute most strongly to the model’s predictions.
Which layer of the neural network does Grad-CAM typically operate on?
The input layer.
The output layer.
The convolutional layers.
The fully connected layers.
Correct answer
The convolutional layers.
Which component of Grad-CAM is responsible for weighting the importance of each feature map?
The class activation map (CAM).
The global average pooling layer.
The guided backpropagation algorithm.
The gradient of the target class score with respect to the feature maps.
Correct answer
The global average pooling layer.
What is the primary role of the Gram matrix in neural style transfer?
To compute the style loss by measuring the similarity between feature maps from the content and style images.
To generate high-resolution images by applying a series of convolutions and upsampling operations.
To regularize the optimization process by penalizing large gradients in the feature space.
To visualize the internal representations learned by the neural network during training.
Correct answer
To compute the style loss by measuring the similarity between feature maps from the content and style images.
What is a potential limitation of non-maximum suppression (NMS) algorithms in object detection?
They are computationally expensive and can significantly increase inference time.
They may fail to suppress redundant bounding boxes in cluttered scenes with multiple overlapping objects.
They require extensive hyperparameter tuning to achieve optimal performance across different datasets.
They are highly sensitive to variations in lighting conditions and viewpoints in the input images.
Correct answer
They may fail to suppress redundant bounding boxes in cluttered scenes with multiple overlapping objects.
What is a key feature of focal loss that distinguishes it from traditional loss functions in object detection?
Focal loss introduces additional regularization terms to penalize large weights in the neural network.
Focal loss dynamically adjusts the loss weights based on the difficulty of each training example.
Focal loss computes the loss based on the Euclidean distance between predicted and groundtruth bounding boxes.
Focal loss prioritizes minimizing the overall loss across all classes equally.
Correct answer
Focal loss dynamically adjusts the loss weights based on the difficulty of each training example.
What is a unique challenge faced by Siamese networks in comparison to traditional neural networks?
Siamese networks require significantly larger datasets for training due to their complex architecture.
Siamese networks need to learn a similarity metric between pairs of input samples, which requires specialized loss functions and training procedures.
Siamese networks are computationally more intensive, making them less scalable for real-time applications.
Siamese networks tend to suffer from overfitting more frequently than traditional neural networks due to their increased model complexity.
Correct answer
Siamese networks need to learn a similarity metric between pairs of input samples, which requires specialized loss functions and training procedures.
Consider the following assertion and reason, and state which of them is true.
Assertion: ProgressiveGAN uses Minibatch Standard Deviation.
Reason: This value will be different if the generated images do not have the same diversity as the actual images, and the discriminator will penalize it as a result.
Both Assertion and Reason are True; Reason is a correct explanation of the Assertion
Both Assertion and Reason are True; but Reason is not a correct explanation of the Assertion
Assertion is True; Reason is False
Assertion is False by itself
Correct answer
Both Assertion and Reason are True; Reason is a correct explanation of the Assertion
Correct answer
Which of the following statements is false?
Reparametrization trick is used in training VAEs to overcome the problem of non-differentiability of the objective function across a sampling step.
In denoising autoencoder, reconstructed output is compared against noise- free ground truth data point.
While training a GAN, Generator and Discriminator weights are not updated concurrently but updated sequentially.
When a GAN is completely trained, Discriminator D outputs very high probability (approximately 1) for images from Generator G, that is, D(G(z)) ≈ 1
Correct answer
When a GAN is completely trained, Discriminator D outputs very high probability (approximately 1) for images from Generator G, that is, D(G(z)) ≈ 1
Match the following:
| 1)VAE | i) |
|---|---|
| 2)GAN | ii)Maximum Likelihood Estimation (MLE) |
| 3)Reparametrization | iii) |
| 4)Generative models | iv) |
Correct answer
Continual learning
Zero-shot learning
Few-shot learning
Unsupervised learning
Correct answer
Few-shot learning
[1.1, 2.06, 1.08]
[0.25, 0.34, 0.18]
[0.2, 0.3, 0.1]
[1.2, 2.3, 1.1]
Correct answer
[0.25, 0.34, 0.18]
What is the advantage of the progressive training methodology proposed in the Progressive GAN paper?
It allows for generating high-resolution images with unlimited variation.
It eliminates the need for using mini-batches during training.
It enables the discovery of large-scale structures before finer-scale details.
It ensures all layers in the networks remain fixed throughout the training process.
Correct answer
It enables the discovery of large-scale structures before finer-scale details.
Contrastive Language Image Pre-training (CLIP) Can be used as a zero-shot classifier.
TRUE
FALSE
Correct answer
TRUE
Maximum likelihood estimate
Minimum likelihood estimate
Maximizing mean square error
Minimizing mean square error
Maximizing negative log likelihood
Minimizing negative log likelihood
Correct answers
Minimum likelihood estimate
Maximizing mean square error
Minimizing mean square error
Maximizing negative log likelihood
The cornerness score of the center pixel for the image represented by the matrix below is __________.
| 1 | 0 | 0 | 0 | 1 |
|---|---|---|---|---|
| 0 | 1 | 0 | 1 | 0 |
| 0 | 0 | 1 | 0 | 0 |
| 0 | 1 | 0 | 1 | 0 |
| 1 | 0 | 0 | 0 | 1 |
NOTE: Enter your answer in two decimal places.
Correct answer: 6.56
Consider a grayscale image in terms of pixel intensities as shown in matrix form below.
| 9 | 1 | 6 |
|---|---|---|
| 3 | 3 | 2 |
| 3 | 5 | 3 |
Assume that each pixel takes values from 1 to 9 (inclusive). After performing histogram equalization, the number of pixels with intensity value 6 are __________?
Correct answer: 4
Correct answer: 4
Consider a Faster R-CNN object detection framework, which is being trained with a mini-batch size of 2. The RPN produces 512 proposals, out of which 10 and 12 of them are foreground proposals, corresponding to each image in the mini-batch. For a single mini-batch update, the bounding box regression loss will be computed _____________ times.
Correct answer: 22
Transformer encoder requires a sequence of __________ as input.
Correct answer: vectors
Given word embeddings for two words and weight matrix to generate queries, keys and values as below (in practice, these would be more complex or learned; for simplicity, we provide them here):
Calculate the query, key, and value to calculate z1. Consider only the first word for calculating the scores; remember that attention output is calculated by a weighted sum of all value vectors. Also, the key vector dimension dk = 3
Based on the above data, answer the given subquestions.
Given word embeddings for two words and weight matrix to generate queries, keys and values as below (in practice, these would be more complex or learned; for simplicity, we provide them here):
First element of attention vector is __________________
Correct answer: 0.969 (accepted within ±0.02)
Given word embeddings for two words and weight matrix to generate queries, keys and values as below (in practice, these would be more complex or learned; for simplicity, we provide them here):
Calculate the query, key, and value to calculate z1. Consider only the first word for calculating the scores; remember that attention output is calculated by a weighted sum of all value vectors. Also, the key vector dimension dk = 3
Based on the above data, answer the given subquestions.
Given word embeddings for two words and weight matrix to generate queries, keys and values as below (in practice, these would be more complex or learned; for simplicity, we provide them here):
Second element of attention vector is _____________
Correct answer: 1 (accepted within ±0.02)
Given word embeddings for two words and weight matrix to generate queries, keys and values as below (in practice, these would be more complex or learned; for simplicity, we provide them here):
Calculate the query, key, and value to calculate z1. Consider only the first word for calculating the scores; remember that attention output is calculated by a weighted sum of all value vectors. Also, the key vector dimension dk = 3
Based on the above data, answer the given subquestions.
Given word embeddings for two words and weight matrix to generate queries, keys and values as below (in practice, these would be more complex or learned; for simplicity, we provide them here):
Third element of attention vector is _________________
Correct answer: 0.03 (accepted within ±0.02)
Based on the above data, answer the given subquestions.
NOTE: Enter your answer in two decimal places.
Correct answer: 3.01
Based on the above data, answer the given subquestions.
NOTE: Enter your answer in two decimal places.
Correct answer: 6.99
Based on the above data, answer the given subquestions.
NOTE: Enter your answer in two decimal places.
Correct answer: 5.01
Based on the above data, answer the given subquestions.
Correct answer: -4
Based on the above data, answer the given subquestions.
Correct answer: -4
Based on the above data, answer the given subquestions.
Correct answer: -4
In vision transformers, if treat each pixel as a token for an image of size 64 × 64 × 3, the number of parameters required for the self-attention calculation are ________blank(a)________ . What will be the number of parameters if consider patches of size 8 × 8 × 3 ________blank(b)________. Based on the above data, answer the given subquestions.
Enter the correct answer for blank(a):
Correct answer: 12288
In vision transformers, if treat each pixel as a token for an image of size 64 × 64 × 3, the number of parameters required for the self-attention calculation are ________blank(a)________ . What will be the number of parameters if consider patches of size 8 × 8 × 3 ________blank(b)________. Based on the above data, answer the given subquestions.
Enter the correct answer for blank(b):
Correct answer: 64
In a ViT, all blocks have same ________blank(a)________ and number of ________blank(b)________ Based on the above data, answer the given subquestions.
Enter the correct answer for blank(a):
Correct answer: resolution
In a ViT, all blocks have same ________blank(a)________ and number of ________blank(b)________ Based on the above data, answer the given subquestions.
Enter the correct answer for blank(b):
Correct answer: channels
