Research map / Detection and segmentation / CNN features
Grid Loss: Detecting Occluded Faces
Michael Opitz, Georg Waltner, Georg Poier, Horst Possegger, Horst Bischof
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 |
|---|---|---|---|
| replace hinge loss with grid loss on output layer | 2-layer CNN detector · FDDB (284 FP) | true positive rate: 0.819 -> 0.851 | Improved |
| add grid loss on hidden layers as auxiliary supervision (ablation) | grid loss + hinge · FDDB (284 FP) | true positive rate: 0.851 -> 0.863 | Improved |
| use block size 2 instead of 16 in grid loss (ablation) | grid loss + hinge on hidden layers · FDDB (284 FP) | true positive rate: 0.834 -> 0.863 | Improved |
| decrease lambda toward zero (down-weight part losses) (ablation) | grid loss · FDDB (284 FP) | true positive rate | Hurt |
| use grid loss instead of hinge loss on heavily occluded faces | 2-layer CNN detector · COFW-HO | true positive rate: 0.909 -> 0.979 | Improved |
| use grid loss instead of hinge loss on less occluded faces | 2-layer CNN detector · COFW-LO | true positive rate: 0.982 -> 0.998 | Improved |
| train with grid loss instead of hinge loss | 2-layer CNN detector · Feature map correlation | sum of off-diagonal correlation: 22500.25 -> 225.96 | Improved |
| train with grid loss on hidden layers when only 1% of positives are used | 2-layer CNN detector · FDDB (284 FP), 0.01 training fraction | true positive rate: 0.700 -> 0.802 | Improved |
| add hinge loss on hidden layers (deep supervision) to baseline | hinge loss CNN · FDDB (284 FP) | true positive rate: 0.819 -> 0.834 | Improved |
| train ellipse regressor with numerical PASCAL overlap instead of SSE (ablation) | ellipse regressor · FDDB continuous evaluation (284 FP) | true positive rate: 0.700 -> 0.702 | No clear effect |
| add ellipse regressor to detector (ablation) | grid loss detector · FDDB (284 FP) | true positive rate: 0.863 -> 0.872 | Improved |
| skip intermediate pyramid scales with ellipse regressor (ablation) | grid loss detector · FDDB (284 FP) | true positive rate: 0.872 -> 0.867 | No clear effect |
Closest papers by meaning
- End2End Occluded Face Recognition by Masking Corrupted Features (TPAMI 2021 · similarity 0.87)
- Increasing CNN Robustness to Occlusions by Reducing Filter Support (ICCV 2017 · similarity 0.86)
- Loss Function Discovery for Object Detection via Convergence-Simulation Driven Search (ICLR 2021 · similarity 0.85)
- Faceness-Net: Face Detection through Deep Facial Part Responses (TPAMI 2017 · similarity 0.85)
- Compositional Convolutional Neural Networks: A Robust and Interpretable Model for Object Recognition Under Occlusion (IJCV 2020 · similarity 0.85)
- G-CNN: An Iterative Grid Based Object Detector (CVPR 2016 · 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.
