Marigold Diffusion Depth Estimation Quiz

Answers and explanations appear as you take the quiz.

1. What pretrained model does Marigold build on?
   • Stable Diffusion
   • GPT-4
   • YOLOv8
   • ResNet-50
2. How does Marigold frame the depth estimation task?
   • As a classification problem
   • As conditional generation of a depth map
   • As reinforcement learning
   • As a clustering problem
3. What was surprising about Marigold's training data?
   • It used billions of real depth scans
   • It needed no training at all
   • It was fine-tuned mainly on synthetic datasets yet generalized to real images
   • It used only audio data
4. Why is Marigold typically slower than feed-forward depth models?
   • It runs on a CPU only
   • It downloads data from the internet each time
   • It re-trains for every image
   • It uses an iterative denoising sampling loop
5. In what space does Marigold perform its diffusion process?
   • Stable Diffusion's VAE latent space
   • Raw pixel space
   • Frequency/Fourier space
   • Point-cloud space
6. How can Marigold improve depth stability across runs?
   • By lowering the image resolution
   • By ensembling multiple sampled predictions
   • By disabling the VAE
   • By converting to grayscale
7. What does affine-invariant relative depth mean for a Marigold prediction?
   • Every pixel gives exact distance in meters without calibration
   • The output contains color labels instead of depth
   • Depth ordering is useful, but the output's absolute scale and offset need not equal real-world measurements
   • Only the nearest object receives a depth value
8. Which pair of inputs does Marigold encode with the same VAE during its latent diffusion process?
   • The camera pose and a 3D mesh
   • The text prompt and a CLIP score
   • Two stereo images from separate cameras
   • The input image and the depth map