Guide-linked quiz · Hard Level
Learned Feature Matching: SuperGlue and LightGlue Quiz
Tests how attention-based matchers reason over keypoints, how optimal transport handles unmatched points, and why LightGlue can stop early on easier pairs.
Question 1 of 8
According to the guide, what is the main limitation of classical nearest-neighbor descriptor matching that SuperGlue addresses?
Preview every question8 questions
Answers and explanations appear as you take the quiz.
- According to the guide, what is the main limitation of classical nearest-neighbor descriptor matching that SuperGlue addresses?
- It cannot run on a CPU
- It requires color images only
- It judges each keypoint in isolation, which struggles with repeated patterns or large viewpoint change
- It cannot detect any keypoints at all
- Who is credited in the guide with introducing SuperGlue, and at which venue?
- Sarlin and colleagues at CVPR in 2020
- Rublee and colleagues at OpenCV
- David Lowe at ICCV
- Zhang and colleagues at ECCV
- What does the guide say SuperGlue uses to build contextual understanding of the whole scene rather than judging keypoints in isolation?
- Convolutional pooling layers only
- A simple k-means clustering step
- A checkerboard calibration pattern
- Alternating self-attention and cross-attention layers in a graph neural network
- What algorithm does SuperGlue use to solve the final matching assignment, as described in the guide?
- RANSAC
- The Sinkhorn algorithm, framing matching as optimal transport
- The eight-point algorithm
- Gradient descent on pixel intensities
- How does SuperGlue's matching formulation handle a keypoint that is visible in only one of the two images, per the guide?
- It forces an incorrect match regardless
- It deletes the keypoint from both images
- It allows the point to be marked as unmatched, naturally handling occlusion
- It always assigns it to the nearest image border
- Which mistaken assumption about SuperGlue and LightGlue does the guide correct?
- That they detect keypoints themselves, when they actually match keypoints found by a separate detector like SuperPoint
- That they run only on mobile devices
- That they cannot be used for 3D reconstruction
- That they require no training data
- According to the technical insight, how does LightGlue achieve faster inference than SuperGlue?
- By skipping attention layers entirely
- By reducing the image resolution before matching
- By avoiding the use of any positional encoding
- By using a confidence-based early-exit mechanism that stops processing once predicted matches are stable
- What does the guide's future outlook say might happen to the separation between detectors like SuperPoint and matchers like LightGlue?
- It will become more rigid and formally standardized
- It may blur further as detectors and matchers get jointly optimized or replaced by end-to-end pipelines
- It will be eliminated by a return to purely classical methods
- Detectors will be phased out entirely with no replacement
Keep testing yourself
More quizzes picked for your level.