Bounding box handling in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of bounding box area calculation

What is the output of the following code that calculates the area of bounding boxes stored as (x_min, y_min, x_max, y_max)?

PyTorch

import torch
boxes = torch.tensor([[1, 2, 4, 6], [0, 0, 3, 3], [2, 2, 5, 5]])
widths = boxes[:, 2] - boxes[:, 0]
heights = boxes[:, 3] - boxes[:, 1]
areas = widths * heights
print(areas.tolist())

A [15, 9, 12]

B [12, 9, 9]

C [12, 6, 9]

D [9, 9, 9]

💡 Hint

Remember area = width * height, where width = x_max - x_min and height = y_max - y_min.

The widths are [3, 3, 3] and heights are [4, 3, 3]. Multiplying element-wise gives areas [12, 9, 9].

❓ Model Choice

intermediate

Choosing the right bounding box format for IoU calculation

Which bounding box format is best suited for calculating Intersection over Union (IoU) directly without conversion?

A Top-left and bottom-right corners (x_min, y_min, x_max, y_max)

B Top-left corner with width and height (x_min, y_min, w, h)

C Center coordinates with corner points (cx, cy, x_min, y_min)

D Center coordinates with width and height (cx, cy, w, h)

💡 Hint

IoU requires overlapping area calculation which is easier with corner coordinates.

The (x_min, y_min, x_max, y_max) format directly gives the corners needed to find intersection areas without extra conversion.

❓ Hyperparameter

advanced

Effect of IoU threshold on Non-Maximum Suppression (NMS)

What is the effect of increasing the IoU threshold parameter in Non-Maximum Suppression when filtering bounding boxes?

A Boxes are sorted by confidence score instead of IoU

B Fewer boxes are kept, resulting in stricter suppression

C No change in number of boxes kept

D More boxes are kept, allowing more overlapping detections

💡 Hint

Higher IoU threshold means boxes must overlap more to be suppressed.

Increasing IoU threshold means boxes with less overlap are not suppressed, so more boxes remain after NMS.

🔧 Debug

advanced

Debugging incorrect bounding box clipping

Given bounding boxes and image size, which option correctly clips boxes to image boundaries without errors?

PyTorch

import torch
boxes = torch.tensor([[50, 30, 200, 180], [-10, 20, 100, 150], [30, 40, 500, 400]])
image_size = (300, 300)  # height, width
# Clip boxes so coordinates stay within image boundaries

A

boxes[:, 0] = boxes[:, 0].clamp(0, image_size[1])
boxes[:, 1] = boxes[:, 1].clamp(0, image_size[0])
boxes[:, 2] = boxes[:, 2].clamp(0, image_size[1])
boxes[:, 3] = boxes[:, 3].clamp(0, image_size[0])

B

boxes[:, 0] = boxes[:, 0].clamp(0, image_size[0])
boxes[:, 1] = boxes[:, 1].clamp(0, image_size[1])
boxes[:, 2] = boxes[:, 2].clamp(0, image_size[0])
boxes[:, 3] = boxes[:, 3].clamp(0, image_size[1])

C

boxes[:, 0] = boxes[:, 0].clamp(0, image_size[1]-1)
boxes[:, 1] = boxes[:, 1].clamp(0, image_size[0]-1)
boxes[:, 2] = boxes[:, 2].clamp(0, image_size[1]-1)
boxes[:, 3] = boxes[:, 3].clamp(0, image_size[0]-1)

D

boxes[:, 0] = boxes[:, 0].clamp(0, image_size[0]-1)
boxes[:, 1] = boxes[:, 1].clamp(0, image_size[0]-1)
boxes[:, 2] = boxes[:, 2].clamp(0, image_size[1]-1)
boxes[:, 3] = boxes[:, 3].clamp(0, image_size[1]-1)

💡 Hint

Remember image_size is (height, width). x coordinates clamp to width, y coordinates clamp to height.

x coordinates (0 and 2) clamp to width (image_size[1]-1), y coordinates (1 and 3) clamp to height (image_size[0]-1).

🧠 Conceptual

expert

Understanding bounding box regression targets in object detection

In object detection, bounding box regression predicts offsets relative to anchor boxes. Which statement correctly describes why offsets are predicted instead of absolute coordinates?

A Offsets normalize the prediction space, making training more stable and easier to learn

B Absolute coordinates are always better because they are direct and simpler

C Offsets reduce the number of bounding boxes needed during inference

D Absolute coordinates allow the model to ignore anchor boxes completely

💡 Hint

Think about how predicting relative changes helps the model focus on small adjustments.

Predicting offsets relative to anchors normalizes the scale and position differences, improving training stability and convergence.