YOLO concept in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

What is the main advantage of YOLO compared to traditional object detection methods?

YOLO (You Only Look Once) is a popular object detection model. What is its main advantage compared to older methods like sliding window or region proposal based detectors?

A It processes the entire image in a single pass, making detection faster.

B It uses multiple passes over the image to improve accuracy.

C It requires manual feature engineering for each object class.

D It detects objects by scanning small patches independently.

💡 Hint

Think about how YOLO treats the image differently from methods that look at many small parts separately.

YOLO divides the image into a grid and predicts bounding boxes and class probabilities for all grid cells in one forward pass, making it very fast compared to methods that scan many regions separately.

❓ Predict Output

intermediate

What is the output shape of YOLO's final layer for a 7x7 grid with 2 boxes per cell and 20 classes?

YOLO divides the image into an SxS grid. Each grid cell predicts B bounding boxes and C class probabilities. Given S=7, B=2, and C=20, what is the shape of the output tensor from the final layer?

A [7, 7, 30]

B [7, 7, 14]

C [7, 7, 24]

D [7, 7, 28]

💡 Hint

Each bounding box predicts 5 values (x, y, w, h, confidence) plus class probabilities per cell.

Each cell predicts B=2 boxes, each with 5 values, so 2*5=10 values, plus C=20 class probabilities, total 30 values per cell. So output shape is [7,7,30].

❓ Model Choice

advanced

Which PyTorch layer is best suited to implement YOLO's final prediction layer?

YOLO's final layer outputs a tensor with shape [batch_size, S, S, B*5 + C]. Which PyTorch layer is most appropriate to produce this output from a feature map?

A nn.Conv2d with kernel size 3 and output channels = B*5 + C

B nn.Conv2d with kernel size 1 and output channels = B*5 + C

C nn.Linear with output features = S*S*(B*5 + C)

D nn.MaxPool2d with kernel size 2

💡 Hint

The layer should keep spatial dimensions but change channel depth to prediction size.

A 1x1 convolution changes the number of channels per spatial location without changing height and width, perfect for predicting bounding boxes and classes per grid cell.

❓ Metrics

advanced

Which metric best evaluates YOLO's object detection performance?

YOLO predicts bounding boxes and class labels. Which metric is most appropriate to evaluate its detection accuracy?

A Accuracy

B Mean Squared Error (MSE)

C Mean Average Precision (mAP)

D Cross-Entropy Loss

💡 Hint

Consider a metric that accounts for both localization and classification correctness.

mAP measures how well predicted boxes match ground truth boxes and correct class labels, combining localization and classification performance.

🔧 Debug

expert

Why does YOLO sometimes miss small objects in images?

YOLO divides the image into a grid and predicts boxes per cell. Why might it struggle to detect very small objects?

A YOLO only predicts one bounding box per image.

B YOLO uses too many grid cells, causing overfitting on small objects.

C YOLO's loss function ignores small objects during training.

D Small objects may fall into one grid cell and be overshadowed by larger objects, causing missed detections.

💡 Hint

Think about how grid size affects the ability to localize small objects.

Because YOLO predicts a fixed number of boxes per grid cell, small objects that occupy less than a cell can be missed or confused with larger objects in the same cell.