Image Captioning Quiz

Answers and explanations appear as you take the quiz.

1. What does an image captioning system produce as output?
   • A natural-language description of the image
   • A higher-resolution version of the image
   • A list of pixel coordinates only
   • A compressed file of the image
2. In a classic captioning pipeline, what role does the visual encoder play?
   • It writes the final sentence word by word
   • It extracts feature representations from the image
   • It translates captions between languages
   • It scores the caption's grammar
3. Why is attention useful in image captioning?
   • It compresses the image to save storage
   • It removes the need for any training data
   • It lets the model focus on relevant image regions for each word
   • It converts the caption into audio
4. Which dataset is widely used for training and evaluating image captioning?
   • MNIST
   • WikiText
   • LibriSpeech
   • MS COCO
5. What does it mean for a caption decoder to generate words 'autoregressively'?
   • It predicts each word using the image and the words generated so far
   • It produces all words simultaneously
   • It only outputs a single word per image
   • It regresses the image back to pixels
6. A common failure mode of captioning models known as 'hallucination' refers to what?
   • Producing captions that are too short
   • Describing objects that are not actually in the image
   • Translating captions incorrectly
   • Running too slowly on large images
7. How many human captions were collected for each image in the original COCO Captions dataset?
   • One caption total for the entire dataset
   • No text descriptions were collected
   • Five captions per image
   • One hundred captions for every image
8. What does CIDEr compare when evaluating an image caption?
   • Only the image’s pixel resolution
   • The caption’s audio waveform against a transcript
   • Whether the caption contains a certain number of words
   • The candidate caption against human reference captions using consensus-weighted n-grams