Research map / Detection and segmentation / CNN features
Uncertainty Quantification in Detection Transformers: Object-Level Calibration and Image-Level Reliability
Young‐Jin Park, Carson Sobolewski, Navid Azizan
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 negative-prediction confidence with weight lambda (contrast) to positive confidence (ablation) | Conf+ (OCE) · COCO, Cal-DETR | Pearson correlation with image reliability: 0.547 -> 0.628 | Improved |
| add contrast with negatives (ablation) | Conf+ (OCE) · COCO, UP-DETR | Pearson correlation with image reliability: 0.477 -> 0.656 | Improved |
| add contrast with negatives (ablation) | Conf+ (OCE) · Foggy Cityscapes, D-DETR | Pearson correlation with image reliability: 0.276 -> 0.350 | Improved |
| select positives with OCE-minimizing post-processing instead of threshold 0.3 | Conf+ (0.3) · Foggy Cityscapes, UP-DETR | Pearson correlation with image reliability: 0.195 -> 0.191 | No clear effect |
| use optimal matching instead of OCE-based separation | ContrastiveConf · COCO, UP-DETR | Pearson correlation with image reliability: 0.656 -> 0.700 | Improved |
| use top-100 predictions as positives | Conf+ · COCO, UP-DETR | Pearson correlation with image reliability: 0.476 -> -0.619 | Hurt |
| use lambda between 1.0 and 10.0 instead of 0.0 in ContrastiveConf (ablation) | ContrastiveConf (OCE) · COCO | Pearson correlation with image reliability | Improved |
| use excessively large lambda (ablation) | ContrastiveConf (OCE) · Foggy Cityscapes | Pearson correlation with image reliability | Hurt |
| replace averaging with Max-IoU or IoU-weighted aggregation in OCE (ablation) | OCE · 6 models x 3 distributions | correlation between OCE variants | No clear effect |
| vary IoU threshold tau in OCE (ablation) | OCE · 6 models x 3 distributions | correlation between OCE variants | No clear effect |
Closest papers by meaning
- Cal-DETR: Calibrated Detection Transformer (NeurIPS 2023 · similarity 0.92)
- Systematic Evaluation of Uncertainty Calibration in Pretrained Object Detectors (IJCV 2024 · similarity 0.90)
- DETRs Beat YOLOs on Real-time Object Detection (CVPR 2024 · similarity 0.89)
- "On Calibration of Object Detectors: Pitfalls, Evaluation and Baselines" (ECCV 2024 · similarity 0.89)
- Towards Building Self-Aware Object Detectors via Reliable Uncertainty Quantification and Calibration (CVPR 2023 · similarity 0.88)
- End-to-End Object Detection with Transformers (ECCV 2020 · 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.
