Research map / Label-efficient and robust learning / Recognition and long-tail
FaceXFormer: A Unified Transformer for Facial Analysis
Kartik Narayan, Vibashan VS, Rama Chellappa, Vishal M. Patel
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MLP fusion of multi-scale features (ablation) | FaceXFormer · BIWI head pose | MAE: 3.70 -> 3.52 | Improved |
| add MLP fusion of multi-scale features (ablation) | FaceXFormer · CelebA attributes | accuracy: 91.21 -> 91.83 | Improved |
| add MLP fusion of multi-scale features (ablation) | FaceXFormer · UTKFace age | MAE: 4.21 -> 4.17 | Improved |
| add MLP fusion of multi-scale features (ablation) | FaceXFormer · 300W landmarks | NME: 4.70 -> 4.67 | Improved |
| add cross-attention between face and task tokens in decoder (ablation) | FaceXFormer · BIWI head pose | MAE: 17.16 -> 4.63 | Improved |
| add cross-attention between face and task tokens in decoder (ablation) | FaceXFormer · 300W landmarks | NME: 31.49 -> 5.30 | Improved |
| replace standard cross-attention with bi-directional cross-attention (ablation) | FaceXFormer · BIWI head pose | MAE: improvements of 1.11 MAE in head pose estimation | Improved |
| replace standard cross-attention with bi-directional cross-attention (ablation) | FaceXFormer · 300W landmarks | NME: 0.63 NME in landmark detection | Improved |
| replace standard cross-attention with bi-directional cross-attention (ablation) | FaceXFormer · CelebA attributes | accuracy: 89.98 -> 91.83 | Improved |
| replace standard cross-attention with bi-directional cross-attention (ablation) | FaceXFormer · UTKFace age | MAE: 0.67 MAE in age estimation | Improved |
| use Swin-B backbone instead of MobileNet (ablation) | FaceXFormer · Backbone comparison (regression tasks) | mean error: 4.64 -> 4.12 | Improved |
| use ConvNeXt-B instead of Swin-B backbone (ablation) | FaceXFormer · Backbone comparison (segmentation) | F1: 92.01 -> 92.08 | Mixed |
Closest papers by meaning
- Faceptor: A Generalist Model for Face Perception (ECCV 2024 · similarity 0.91)
- Pre-training Strategies and Datasets for Facial Representation Learning (ECCV 2022 · similarity 0.86)
- DeepCoder: Semi-Parametric Variational Autoencoders for Automatic Facial Action Coding (ICCV 2017 · similarity 0.85)
- FSRT: Facial Scene Representation Transformer for Face Reenactment from Factorized Appearance Head-pose and Facial Expression Features (CVPR 2024 · similarity 0.85)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.84)
- DSL-FIQA: Assessing Facial Image Quality via Dual-Set Degradation Learning and Landmark-Guided Transformer (CVPR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
