Research map / Label-efficient and robust learning / Recognition and long-tail
DeCaFA: Deep Convolutional Cascade for Face Alignment in the Wild
Arnaud Dapogny, Kevin Bailly, Matthieu Cord
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 |
|---|---|---|---|
| reinject whole input image (F3 instead of F2 fusion) (ablation) | DeCaFA · 300W-Challenging | mean error (%): 5.45 -> 5.38 | Improved |
| reinject whole input image (F3 instead of F2 fusion) (ablation) | DeCaFA · WFLW-Pose | mean error (%): 8.78 -> 8.39 | Improved |
| use F5 fusion instead of basic deep F1 fusion (ablation) | DeCaFA · Average over 5 subsets | mean error (%): 4.83 -> 4.69 | Improved |
| use F4 fusion instead of F1 fusion (ablation) | DeCaFA · Average over 5 subsets | mean error (%): 4.83 -> 4.80 | No clear effect |
| fuse with attention maps (F2 instead of F1) (ablation) | DeCaFA · Average over 5 subsets | mean error (%): 4.83 -> 5.04 | Hurt |
| chain transfer layers instead of independent ones (ablation) | DeCaFA · Average over 5 subsets | mean error (%): 4.75 -> 4.69 | Improved |
| increasing intermediate supervision weights instead of constant (ablation) | DeCaFA · Average over 5 subsets | mean error (%): 4.91 -> 4.69 | Improved |
| increasing intermediate supervision weights instead of decreasing (ablation) | DeCaFA · Average over 5 subsets | mean error (%): 5.05 -> 4.69 | Improved |
| increasing intermediate supervision weights instead of decreasing (ablation) | DeCaFA · WFLW-Pose | mean error (%): 9.06 -> 8.11 | Improved |
| add CelebA and WFLW auxiliary data to 300W training (ablation) | DeCaFA · 300W (full) | mean error (%): 3.69 -> 3.39 | Improved |
| add 300W and CelebA auxiliary data to WFLW training (ablation) | DeCaFA · WFLW all | mean error (%): 5.01 -> 4.62 | Improved |
| add coarsely annotated CelebA (5 landmarks) data when training on few fine-annotated images (ablation) | DeCaFA · 300W and WFLW with 3%/15% and 1%/6% of train set | mean error (%) | Improved |
Closest papers by meaning
- Robust Face Alignment via Deep Progressive Reinitialization and Adaptive Error-driven Learning (TPAMI 2021 · similarity 0.89)
- Pose-Invariant Face Alignment With a Single CNN (ICCV 2017 · similarity 0.89)
- Mnemonic Descent Method: A Recurrent Process Applied for End-To-End Face Alignment (CVPR 2016 · similarity 0.89)
- Recurrent Convolutional Shape Regression (TPAMI 2018 · similarity 0.89)
- Direct Shape Regression Networks for End-to-End Face Alignment (CVPR 2018 · similarity 0.88)
- Self-Reinforced Cascaded Regression for Face Alignment (AAAI 2018 · similarity 0.88)
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.
