Research map / Detection and segmentation / CNN features
Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection
Nishant Kumar, Siniša Šegvić, Abouzar Eslami, Stefan Gumhold
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 |
|---|---|---|---|
| Use a joint normalizing-flow distribution and synthesize low-likelihood outlier features instead of class-conditional Gaussian synthesis | FFS object detector · PASCAL-VOC → MS-COCO | FPR95: 47.77 -> 44.15 | Improved |
| Use a joint normalizing-flow distribution and synthesize low-likelihood outlier features instead of class-conditional Gaussian synthesis | FFS object detector · PASCAL-VOC → OpenImages | AUROC: 87.59 -> 88.29 | Improved |
| Use flow-based feature synthesis rather than STUD | FFS object detector · BDD100K → nuImages | mAP: 32.3 -> 36.2 | Improved |
| Use flow-based feature synthesis rather than STUD | FFS object detector · Youtube-VIS → MS-COCO | FPR95: 81.14 -> 83.06 | Hurt |
| Add rejection sampling to VOS outliers so samples unlikely under any class Gaussian are retained (ablation) | VOS · PASCAL-VOC → MS-COCO | FPR95: 48.22 -> 45.59 | Improved |
| Replace Glow with NICE (ablation) | FFS · PASCAL-VOC → MS-COCO | FPR95: 44.15 -> 48.23 | Hurt |
| Replace Glow with RealNVP (ablation) | FFS · PASCAL-VOC → MS-COCO | FPR95: 44.15 -> 45.76 | Hurt |
| Replace Glow with GIN (ablation) | FFS · PASCAL-VOC → MS-COCO | AUROC: 89.71 -> 88.36 | Hurt |
| Replace BCE regularization with CE (ablation) | FFS · PASCAL-VOC → MS-COCO | FPR95: 44.15 -> 69.92 | Hurt |
| Replace BCE regularization with JSD (ablation) | FFS · PASCAL-VOC → MS-COCO | FPR95: 44.15 -> 52.83 | Hurt |
| Replace BCE regularization with Hinge (ablation) | FFS · PASCAL-VOC → MS-COCO | FPR95: 44.15 -> 51.32 | Hurt |
| Increase the number of sampled projection outliers beyond a moderate amount (ablation) | FFS projection sampling · PASCAL-VOC → MS-COCO | FPR95: 47.23 -> 59.60 | Hurt |
Closest papers by meaning
- Understanding Likelihood of Normalizing Flow and Image Complexity through the Lens of Out-of-Distribution Detection (AAAI 2024 · similarity 0.88)
- Towards Open-World Object-based Anomaly Detection via Self-Supervised Outlier Synthesis (ECCV 2024 · similarity 0.88)
- SAFE: Sensitivity-Aware Features for Out-of-Distribution Object Detection (ICCV 2023 · similarity 0.87)
- RODEO: Robust Outlier Detection via Exposing Adaptive Out-of-Distribution Samples (ICML 2024 · similarity 0.85)
- Quantile-Based Maximum Likelihood Training for Outlier Detection (AAAI 2024 · similarity 0.85)
- Robustly Train Normalizing Flows via KL Divergence Regularization (AAAI 2024 · 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.
