Skip to content
← Models
Open weightsOtto von Guericke University Magdeburg/10 trustMetrics-checked

L2CS-Net

Fine-grained CNN gaze estimation for unconstrained settings — separate angle regression for sharper eye tracking.

Visit official page →

About

L2CS-Net estimates 3D gaze direction from face images in the wild, and its trick is disarmingly simple: instead of regressing the gaze vector jointly, it regresses the yaw and pitch angles separately with dedicated losses for each. That per-angle treatment squeezes out finer accuracy, and it landed state-of-the-art-at-release numbers — 3.92° of angular error on MPIIGaze and 10.41° on the full 180° range of Gaze360.

The code is MIT-licensed and the weights are freely downloadable, which is why it's still a default gaze baseline in toolkits like py-feat (which re-hosts the weights as safetensors with a documented conversion). It's a 2022-era CNN, so it won't win a pure accuracy shootout against newer transformer-based gaze models, but for a permissively licensed, well-documented gaze estimator you can drop into a pipeline today, it holds up.

Specs

Context window

Not applicable — vision model; face crop input

Pricing

Free / open weights

Free quota

Fully open — no API

API access

No official API — self-host from the Ahmednull/L2CS-Net repo, or use it via py-feat

Weights

Open weights

Released

Not disclosed

Benchmarks

  • Paper-reported (arXiv 2203.03339): 3.92° angular error on MPIIGaze
  • Same paper: 10.41° (front 180°) / 9.01° (front-facing) on Gaze360
  • Re-hosted in py-feat with a documented safetensors conversion

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

eye-trackinggaze-estimationcomputer-vision

Trust score — 7/10

/10 trustMetrics-checked

Scored 7/10 (TIER metrics, metrics-v1). Evidence: arXiv paper (2203.03339) with SOTA-at-release numbers on MPIIGaze and Gaze360; MIT-licensed official repo (Ahmednull/L2CS-Net); independently re-hosted in py-feat with a documented model card. Strengths: permissive license, transparent benchmarks, still a strong baseline. Docked for: (1) repo quiet since 2022 — superseded in some settings by transformer-based gaze models; (2) headline numbers are author-reported. License: MIT.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
6.0/10
Maintenance · recency, releases, issue load
4.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · arXiv paper 2203.03339; authors at Otto von Guericke University Magdeburg
  • · Paper reports 3.92 deg error on MPIIGaze and 10.41 deg (front 180) / 9.01 deg (front-facing) on Gaze360
  • · Independently re-hosted in py-feat with a documented safetensors conversion; repo quiet since 2022

Sources