Research map / Label-efficient and robust learning / Domain adaptation and generalization

From Feature to Gaze: A Generalizable Replacement of Linear Layer for Gaze Estimation

CVPR 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Cross-domain few-shot with transformers · 15 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AGG, a domain generalization framework for gaze estimation that replaces the final FC layer with a geodesic-projection (Isomap) plus sphere-alignment module (GPM) and trains the backbone with Sphere-Oriented Training, using only source-domain data.

Problem

Appearance-based gaze estimation models degrade substantially on unseen target domains, partly because the high-dimensional FC regression layer overfits to gaze-irrelevant features; the task requires generalization without any target-domain data.

Approach

Observes that geodesic distances between 512D CNN features are proportional to gaze angle differences, so Isomap projects features to a 3D sphere that is aligned to the gaze unit sphere with only 10 learnable parameters (GPM); Sphere-Oriented Training then optimizes the CNN through an MLP surrogate (Isometric Propagator) of Isomap and the inverse alignment.

Key results

• AGG improves ResNet-18 cross-domain error by up to 35.79% (DG→DD, 12.35 to 7.93 degrees) without target data, improving all 12 reported cross-domain settings.
• With ResNet-18, AGG (e.g., DE→DD 7.07 degrees) outperforms SOTA methods such as PureGaze, ADL and CA-Net in three of four cross-domain settings.
• GPM alone (10 parameters replacing the FC layer) beats the FC layer in 7 of 8 cross-domain experiments in Table 1.

Measured effects reported in the paper

Surprising findings

• Within-dataset accuracy of the GPM and AGG models is sometimes worse, e.g. ResNet-18 within-DG error rises from 12.73 to 13.03 degrees, which the authors attribute to reduced overfitting.
• The 2D Euler-angle gaze representation breaks the sphere structure of the projected features, so the sphere alignment must be redesigned.

Limitations

• Isomap cannot be integrated directly into backpropagation, so an MLP Isometric Propagator approximates it and may itself overfit.
• Only 3D unit-vector gaze representation works; 2D Euler-angle representations break the sphere alignment and need redesign.
• The DE and DG source settings were excluded from evaluation due to large cross-domain error (around 20 degrees).

Experimental setup

Datasets: ETH-XGaze Gaze360 MPIIFaceGaze EyeDiap

Benchmarks: ETH-XGaze Gaze360 MPIIFaceGaze EyeDiap

Metrics: angular gaze error (degrees) sphere error

Baselines: Full-Face ADL CA-Net LatentGaze PureGaze ResNet-18 ResNet-50 VGG16

Builds on: Isomap PureGaze L1 loss pretraining

Closest papers by meaning

• Gaze from Origin: Learning for Generalized Gaze Estimation by Embedding the Gaze Frontalization Process (AAAI 2024 · similarity 0.91)
• Enhancing 3D Gaze Estimation in the Wild using Weak Supervision with Gaze Following Labels (CVPR 2025 · similarity 0.88)
• CLIP-Gaze: Towards General Gaze Estimation via Visual-Linguistic Model (AAAI 2024 · similarity 0.87)
• A Generalized Label Shift Perspective for Cross-Domain Gaze Estimation (NeurIPS 2025 · similarity 0.87)
• Gaze Label Alignment: Alleviating Domain Shift for Gaze Estimation (AAAI 2025 · similarity 0.87)
• De^2Gaze: Deformable and Decoupled Representation Learning for 3D Gaze Estimation (CVPR 2025 · similarity 0.86)