Perceptual Loss and LPIPS Quiz

Answers and explanations appear as you take the quiz.

1. Why does pixel-based L2 loss often misjudge image similarity compared to perceptual loss?
   • It penalizes tiny shifts and texture changes humans barely notice
   • It runs too slowly to be practical
   • It can only compare grayscale images
   • It requires a pretrained network to compute
2. What does LPIPS primarily compare between two images?
   • Raw RGB pixel values
   • File sizes after compression
   • Activations from intermediate layers of a pretrained network
   • Color histograms
3. How were the per-channel weights in LPIPS determined?
   • Set to all-ones by default
   • Calibrated against human perceptual similarity judgments
   • Hand-tuned by the authors
   • Derived from JPEG compression tables
4. What does a lower LPIPS value indicate for a tested image pair?
   • Proof that the two images have identical pixels
   • A guarantee that every human will prefer the lower-scoring image
   • A measurement of the images' compressed file size
   • A smaller distance under the chosen learned perceptual metric
5. Which backbone network is most commonly associated with classic perceptual (feature) loss?
   • ResNet-152
   • YOLO
   • VGG
   • U-Net
6. Which task most directly benefits from using perceptual loss instead of pure L2?
   • Photo-realistic super-resolution
   • Counting pixels in an image
   • Reading EXIF metadata
   • Sorting images by filename
7. What image-input convention does the official LPIPS Python example specify?
   • Raw compressed JPEG bytes used directly as feature maps
   • Only filename strings with no decoded image data
   • A single grayscale class label for each image
   • RGB tensors with image values scaled to the range minus one to plus one
8. Which differences receive learned per-channel weighting inside LPIPS?
   • Differences between the images' filename lengths
   • Squared differences between normalized deep feature representations
   • Only differences between JPEG quality settings
   • Only the number of object class labels predicted