Faster R-CNN usage in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output shape of Faster R-CNN model predictions

Given the following PyTorch code snippet using a pretrained Faster R-CNN model, what is the type and structure of the output after running the model on a batch of images?

PyTorch

import torch
from torchvision.models.detection import fasterrcnn_resnet50_fpn

model = fasterrcnn_resnet50_fpn(pretrained=True)
model.eval()

images = [torch.rand(3, 300, 400), torch.rand(3, 500, 600)]
outputs = model(images)

print(type(outputs))
print(len(outputs))
print(outputs[0].keys())

A outputs is a list of length 2 (one per image), each element is a dict with keys 'boxes', 'labels', and 'scores'.

B outputs is a tensor of shape (2, 100, 6) containing bounding boxes and labels for all images.

C outputs is a dict with keys 'boxes', 'labels', and 'scores' containing tensors for all images combined.

D outputs is a list of length 1 containing a dict with keys 'boxes', 'labels', and 'scores'.

💡 Hint

Think about how torchvision detection models return predictions for each image separately.

The Faster R-CNN model returns a list of dictionaries, one per input image. Each dictionary contains keys 'boxes', 'labels', and 'scores' with tensors describing detected objects.

❓ Model Choice

intermediate

Choosing the right backbone for Faster R-CNN

You want to detect small objects in high-resolution images using Faster R-CNN. Which backbone architecture is generally better for capturing fine details?

A ResNet-18 without FPN

B ResNet-50 with Feature Pyramid Network (FPN)

C MobileNetV2 without FPN

D VGG16 without FPN

💡 Hint

Feature Pyramid Network helps detect objects at multiple scales.

ResNet-50 with FPN provides strong multi-scale feature maps, which help detect small objects better than shallower or backbone without FPN.

❓ Hyperparameter

advanced

Adjusting Faster R-CNN training for fewer false positives

During training Faster R-CNN, you notice many false positive detections. Which hyperparameter adjustment is most likely to reduce false positives?

A Increase the batch size to 64

B Decrease the number of proposals generated by RPN

C Increase the score threshold for predicted boxes during inference

D Lower the learning rate drastically

💡 Hint

False positives can be filtered by confidence score thresholds.

Increasing the score threshold during inference filters out low-confidence boxes, reducing false positives.

🔧 Debug

advanced

Identifying error in Faster R-CNN input format

You run this code and get a runtime error: import torch from torchvision.models.detection import fasterrcnn_resnet50_fpn model = fasterrcnn_resnet50_fpn(pretrained=True) model.eval() image = torch.rand(3, 224, 224) output = model(image) What is the cause of the error?

A The pretrained model cannot run in eval mode.

B The image tensor must have shape (224, 224, 3) instead of (3, 224, 224).

C The model requires images to be normalized before input.

D The model expects a list of images, not a single tensor.

💡 Hint

Check the expected input type for torchvision detection models.

Faster R-CNN expects a list of tensors, each representing an image. Passing a single tensor causes a runtime error.

❓ Metrics

expert

Evaluating Faster R-CNN with mAP metric

You trained a Faster R-CNN model and want to evaluate its performance using mean Average Precision (mAP) at IoU=0.5. Which statement about mAP is correct?

A mAP@0.5 measures the average precision of detected boxes that have Intersection over Union (IoU) greater than 0.5 with ground truth boxes.

B mAP@0.5 is the average recall of all predicted boxes with confidence scores above 0.5.

C mAP@0.5 counts all predicted boxes regardless of IoU and averages their confidence scores.

D mAP@0.5 is the accuracy of classifying images as containing objects or not.

💡 Hint

mAP involves precision and IoU thresholds for matching predictions to ground truth.

mAP@0.5 calculates average precision considering only predicted boxes with IoU > 0.5 to ground truth, reflecting detection quality.