Optical Flow and Local Feature Matching
Learning where each pixel or keypoint in one image moves to in another image, the dense and sparse correspondence that video, mapping and 3D pipelines depend on.
Tier Grades
Tier A: Frontier AI labs
feasible
Frontier labs can run this work easily, though the exemplars come from universities and Magic Leap rather than frontier labs.
Tier B: Indonesian AI labs
feasible
RAFT came from a 2-author team, a size an Indonesian lab can field for new correspondence models.
Tier C: Institution units such as ITB
feasible
SuperPoint trains on the public MS-COCO set and RAFT on public flow benchmarks, so a university unit can reproduce and extend them.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Released matchers such as LightGlue are meant for latency-sensitive 3D reconstruction, which suits survey and inspection proofs of concept.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeTraining compute is not in the sources read; the RAFT paper states only that it is efficient in inference time, training speed and parameter count.
- DataSuperPoint is trained on the MS-COCO generic image dataset with Homographic Adaptation; RAFT reports a KITTI F1-all error of 5.10 percent and a Sintel final-pass end-point error of 2.855 pixels.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 2 authors on the RAFT paper (arXiv 2003.12039) and 3 on LightGlue (arXiv 2306.13643).
- Time horizonMonths; small teams publish new matchers and flow networks within one conference cycle.
- Shelf-life signalsFeed-forward 3D models that output matches and tracks directly replace separate flow and matching stages, LightGlue adapts its compute to how hard an image pair is, opening latency-sensitive uses
- Pipeline roleTier C and university labs publish flow and matching networks; Tier B adapts them to local imagery; Tier D builds them into mapping, video stabilisation and inspection proofs of concept, and Tier A folds them into feed-forward 3D models.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
LightGluevisionexemplar
LoFTRvisionexemplar
RAFTvisionexemplar
SuperPointvisionexemplar
MASt3R3Dproduced by
ALIKEDvisionproduced by
GMAvisionproduced by
RIFEvisionproduced byMaskFlownetvisionproduced by
IRRvisionproduced by
LiteFlowNetvisionproduced by
vid2depthvisionproduced by
TOFlowvideo generationproduced byPWC-Netvisionproduced by
FGFAvisionproduced byDFF (Deep Feature Flow)visionproduced by
FlowNetvisionproduced byLinks
R&D TopicOptical Flow and Local Feature Matching
ModelsRAFTPrinceton University, visionSuperPointMagic Leap, visionLightGlueETH Zurich, visionLoFTRZhejiang University, vision+13 more
RAFTPrinceton University, visionSuperPointMagic Leap, visionLightGlueETH Zurich, visionLoFTRZhejiang University, vision+13 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 17
DFF (Deep Feature Flow)University of Science and Technology of China (USTC), Microsoft Research, visionproduced byFGFAUniversity of Science and Technology of China (USTC), Microsoft Research, visionproduced byMASt3RNaver, 3Dproduced byRIFEMegvii, Peking University, visionproduced byTOFlowMassachusetts Institute of Technology, Google, video generationproduced byvid2depthUniversity of Texas at Austin, Google, visionproduced byFlowNetUniversity of Freiburg, Technical University of Munich, visionproduced byLoFTRZhejiang University, SenseTime, visionexemplarSuperPointMagic Leap, visionexemplarALIKEDBeihang University, visionproduced byGMAAustralian National University, visionproduced byIRRTU Darmstadt, visionproduced byPapers
- FlowNet: Learning Optical Flow with Convolutional NetworksPhilipp Fischer et al., arXiv preprint, 2015 founding
- SuperPoint: Self-Supervised Interest Point Detection and DescriptionDaniel DeTone et al., CVPR 2018 (per arXiv comment), 2017 founding
- RoMa: Robust Dense Feature MatchingJohan Edstedt et al., arXiv preprint, 2023 state of the art
- LightGlue: Local Feature Matching at Light SpeedPhilipp Lindenberger et al., arXiv preprint, 2023 state of the art
- LoFTR: Detector-Free Local Feature Matching with TransformersJiaming Sun et al., CVPR 2021 (per arXiv comment), 2021 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2003.12039 https://arxiv.org/abs/2003.12039
- arxiv.org/abs/2306.13643 https://arxiv.org/abs/2306.13643
- arxiv.org/abs/1712.07629 https://arxiv.org/abs/1712.07629
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Optical Flow and Local Feature Matching", https://aurion.semai.id/topics/optical-flow-and-local-feature-matching/, records of October 2026.