Learned vs hand-crafted

Classical SLAM is a pipeline of hand-designed modules: a corner detector (FAST, Harris), a descriptor (ORB, SIFT), a matcher (nearest neighbor + ratio test), robust estimation (RANSAC), and a nonlinear least-squares back-end. Every one of these components encodes human assumptions — what a “good” feature looks like, what noise distributions to expect, which heuristics reject outliers. Deep learning offers an alternative: let a network learn these functions from data.

Two very different strategies

There are two ways to apply learning to SLAM, and it is important to keep them apart:

The module-by-module replacement map

Most of Level 5 is organized around which classical module a paper replaces:

Pipeline stageHand-crafted standardLearned replacement
Keypoint detection + descriptionFAST/Harris + ORB/SIFTSuperPoint, R2D2, DISK, XFeat
Matchingnearest neighbor + ratio testSuperGlue, LightGlue, LoFTR (detector-free)
Place recognitionbag-of-words (DBoW)NetVLAD, Patch-NetVLAD, HF-Net
Depthstereo block matching / depth sensorMonoDepth, MiDaS, Depth Anything
Optical flowLucas-Kanade, variational flowRAFT, SEA-RAFT
Relocalizationdescriptor matching against a 3D mapPoseNet (regression), DSAC/ACE (scene coordinates)
Outlier rejection / solverRANSAC, Gauss-NewtonDSAC (soft RANSAC), unrolled/implicit BA layers

Reading a new paper, the first question to ask is: which row is it, and does it keep the classical back-end?

Where each side wins

Learned modules clearly win under conditions that break hand-crafted assumptions: severe illumination change, weak texture, motion blur, seasonal or day-night appearance change, and wide-baseline matching. Learned features are trained on exactly this variation, while a hand-crafted descriptor sees only its designer’s imagination.

Classical modules still win in well-textured, well-lit scenes at high frame rates — ORB is orders of magnitude cheaper to compute; classical geometry gives exact, verifiable solutions with no training-distribution concerns; and failure modes are analyzable. A RANSAC failure can be diagnosed from inlier counts; a network failure is often silent.

End-to-end pose regression specifically loses to geometry: a network that regresses pose directly (PoseNet-style) behaves more like image retrieval than like triangulation, and its errors do not shrink with more texture or better calibration the way geometric methods’ errors do. This is why the field converged on hybrids — learned perception, geometric estimation.

Practical rules of thumb

QuestionLeans classicalLeans learned
Compute budgetEmbedded CPUGPU available
EnvironmentControlled, texturedAppearance change, low texture
Interpretability / certificationRequiredOptional
Training data for your domainScarcePlentiful

Common pitfalls when adopting learned modules

The hybrid consensus

In practice, the most successful modern systems are hybrids: learned perception feeding classical geometric optimization (e.g. learned features inside an ORB-SLAM-style pipeline, or DROID-SLAM’s learned updates around a classical BA solver). The maximum-likelihood machinery of SLAM is hard to beat with a black box; the perception layers that feed it are where learning pays off first. The Differentiability note covers the technology that lets these two halves be trained together.

Why it matters for SLAM

Almost every paper in Level 5 is a point on the learned-vs-hand-crafted spectrum, and knowing where a method sits tells you what problem it solves and what it costs. When designing a system, this framing turns “should I use deep learning?” into concrete engineering questions: which module is failing, whether you can afford inference on your platform, and whether the geometry should remain classical.