Research map / Learning theory and algorithms / Prediction and decision losses
Deep Mixture of Linear Inverse Regressions Applied to Head-Pose Estimation
Stephane Lathuiliere, Remi Juge, Pablo Mesejo, Rafael Munoz-Salinas, Radu Horaud
Check 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 |
|---|---|---|---|
| Replace the Euclidean-loss output layer with a jointly trained Gaussian mixture of linear inverse regressions | VGG-16 head-pose estimator · Biwi | mean absolute error (degrees): 4.34 -> 3.62 | Improved |
| Use inverse regression instead of SVR | HOG features · Biwi | mean absolute error (degrees): 6.66 -> 5.62 | Improved |
| Fine-tune the network using inverse regression | VGG-16 head-pose estimator · Biwi | mean absolute error (degrees): 17.60 -> 3.62 | Improved |
| Use K=1 instead of K=2 Gaussian components (ablation) | VGG-IR-FT · Biwi | pitch MAE (degrees): 4.68 -> 5.12 | Hurt |
| Use K=1 instead of K=2 Gaussian components (ablation) | VGG-IR-FT · S2G and SSV | pitch, yaw, and roll MAE (degrees): S2G: 2.87/3.54/1.65 vs 2.53/3.36/1.97; SSV: 3.12/3.75/2.01 vs 2.86/3.83/2.04 | Mixed |
| Use K=5 instead of K=2 Gaussian components (ablation) | VGG-IR-FT · Biwi | pitch, yaw, and roll MAE (degrees): 4.68/3.12/3.07 -> 5.55/3.74/4.01 | Hurt |
| Use K=10 instead of K=2 Gaussian components (ablation) | VGG-IR-FT · Biwi | pitch, yaw, and roll MAE (degrees): 4.68/3.12/3.07 -> 5.93/3.45/3.99 | Hurt |
| Reduce the training set from 100% to 20% | VGG-IR-FT · Biwi | mean absolute error (degrees): 3.62 -> 7.55 | Hurt |
| Use only RGB rather than additional or depth information | VGG-IR-FT · Biwi | mean absolute error (degrees): 4.23 -> 3.62 | Improved |
| Use K=1 by default rather than a larger mixture (ablation) | VGG-IR-FT · Biwi, S2G, and SSV | pitch, yaw, and roll MAE (degrees): Biwi K=1: 5.12/3.45/3.39; K=2: 4.68/3.12/3.07; K=5: 5.55/3.74/4.01; K=10: 5.93/3.45/3.99 | Mixed |
Closest papers by meaning
- A Comprehensive Analysis of Deep Regression (TPAMI 2020 · similarity 0.89)
- Human Pose Estimation via Convolutional Part Heatmap Regression (ECCV 2016 · similarity 0.86)
- Adversarial Pose Regression Network for Pose-Invariant Face Recognitions (AAAI 2021 · similarity 0.85)
- Mnemonic Descent Method: A Recurrent Process Applied for End-To-End Face Alignment (CVPR 2016 · similarity 0.85)
- Chirality Nets for Human Pose Regression (NeurIPS 2019 · similarity 0.85)
- Adversarial PoseNet: A Structure-Aware Convolutional Network for Human Pose Estimation (ICCV 2017 · similarity 0.85)
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.
