AI computer vision - Part II

Which statement best describes the main difference between a one-stage and a two-stage object detector?

• One-stage detectors use only RGB images, whereas two-stage detectors only work with grayscale images.

• One-stage detectors can detect only small objects, while two-stage detectors only detect large objects.

• One-stage detectors combine region proposal and classification steps into a single network, while two-stage detectors first generate region proposals and then classify them.

• One-stage detectors are unsupervised, and two-stage detectors are fully supervised.

When collecting data for training an object detection model, which is the most important practice to ensure robust performance?

• Only collect images in perfect lighting conditions.

• Make sure all images have the same size, color, and orientation.

• Capture a wide variety of scenarios, viewpoints, and environmental conditions.

• Randomly download images from the internet without confirming relevance.

In the context of object detection, which statement about data annotation is most accurate?

• You only need to provide class labels for entire images; bounding boxes are not necessary.

• Object detection models work best with polygon annotations instead of bounding boxes.

• Bounding boxes must be drawn around each object instance, and each bounding box must be assigned the correct class label.

• Annotation should be done with an automated tool without human review.

Which of the following is not typically used as a data augmentation technique in object detection pipelines?

• Random horizontal flips or rotations of images.

• Scaling or cropping images and corresponding bounding boxes.

• Adjusting brightness, contrast, or color channels.

• Automatically labeling additional training examples without human verification.

During the model development phase for an object detector, which approach is generally best practice?

• Train the model on as many objects as possible at once, regardless of label quality.

• Avoid using pre-trained weights to ensure your model is not biased.

• Avoid using pre-trained weights to ensure your model is not biased.

• Use a well-known base architecture (e.g., ResNet, MobileNet) as a backbone and fine-tune with your labeled dataset.

Which metric is most commonly used to evaluate object detectors on a dataset ?

• Accuracy (percentage of correct classifications).

• Mean Squared Error (MSE).

• Mean Average Precision (mAP) at different IoU thresholds.

• Structural Similarity Index (SSIM).

When deploying an object detection model to an edge device, which is typically the main concern?

• Prioritizing real-time inference with limited compute, memory, and power resources.

• Having the largest possible model to ensure maximum accuracy.

• Limiting the batch size to exactly one image per day.

• Ensuring the model uses random initialization for each inference.