Computer Vision Interview 20 essential Q&A Updated 2026
SLAM

SLAM: 20 Essential Q&A

Simultaneous localization and mapping—poses, landmarks, and closing the loop.

~12 min read 20 questions Advanced
VOloop closurekeyframesbundle
1 What is SLAM? ⚡ easy
Answer: Simultaneous Localization and Mapping—robot/sensor estimates its pose while building a map of the environment from streaming measurements.
2 Localization vs mapping? 📊 medium
Answer: Localization: know map, find pose. Mapping: know poses, build structure. SLAM solves both together when neither is fully known initially.
3 What is visual odometry? 📊 medium
Answer: Estimate incremental camera motion between frames from visual features—local trajectory; drifts without global corrections (part of SLAM front-end).
4 Why keyframes? 📊 medium
Answer: Process only informative frames to limit compute—keyframes anchor map points and graph optimization while skipping redundant views.
5 What is loop closure? 🔥 hard
Answer: Recognize revisiting a place and add constraint to correct accumulated drift—critical for long trajectories.
6 What is drift? ⚡ easy
Answer: Small per-frame errors accumulate so map and pose slowly diverge from truth—loop closure and global BA reduce drift.
7 Pose graph? 📊 medium
Answer: Nodes are keyframe poses, edges are relative measurements (VO, loops)—optimize graph to distribute error (pose-graph relaxation).
8 Bundle adjustment in SLAM? 🔥 hard
Answer: Non-linear least squares jointly refining 3D landmarks and camera poses given reprojection errors—local BA on recent keyframes + global less often.
9 Sparse vs dense SLAM? 📊 medium
Answer: Sparse: point landmarks (features). Dense: surface or depth maps (DTAM, LSD-SLAM)—different compute and representation tradeoffs.
10 Monocular scale? 🔥 hard
Answer: Pure monocular VO/SLAM recovers trajectory up to similarity—absolute scale needs stereo, IMU, or known object size.
11 VIO? 📊 medium
Answer: Fuse camera with IMU for better short-term motion, scale in monocular setups, and robustness to fast motion / brief visual loss.
12 LiDAR SLAM? 📊 medium
Answer: ICP / feature-based registration of 3D scans—accurate metric maps indoors/outdoors; different from pure visual SLAM but same SLAM problem structure.
13 ORB-SLAM family? 📊 medium
Answer: Feature-based pipeline with tracking, local mapping, loop closing threads—popular open-source baseline for monocular/stereo/RGB-D.
14 Relocalization? 📊 medium
Answer: After tracking loss, recover pose by matching current frame to map—database of keyframe descriptors often used.
15 Map formats? ⚡ easy
Answer: Point clouds, surfels, occupancy grids, TSDF volumes—choice depends on navigation, rendering, or planning downstream.
16 Front-end vs back-end? 📊 medium
Answer: Front-end: tracking, data association. Back-end: graph optimization, loop closure—split for modularity and real-time constraints.
17 Data association? 🔥 hard
Answer: Decide if observations correspond to existing landmarks or are new—errors cause map corruption (“ghost” points).
18 Kidnapping problem? 🔥 hard
Answer: Sensor teleported to unrelated place—relocalization must detect inconsistency and reset or search globally.
19 NeRF / learned SLAM? 🔥 hard
Answer: Emerging methods represent scene with neural fields for view synthesis—research intersection with classical SLAM representations.
20 Applications? ⚡ easy
Answer: Robotics navigation, AR persistence, drones, autonomous mapping—any mobile agent needing unknown environment model.

SLAM Cheat Sheet

Core
  • Pose + map
  • VO + back-end
Correct
  • Loop closure
  • BA
Sensors
  • Mono / stereo
  • IMU / LiDAR

💡 Pro tip: Loop closure fights long-term drift.

Full tutorial track

Go deeper with the matching tutorial chapter and code examples.