Research map / Detection and segmentation / Open-set and few-shot detection
Multiclass Confidence and Localization Calibration for Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 MCCL train-time calibration to FCOS | FCOS · In-domain datasets: Sim10K, KITTI, Cityscapes, COCO, VOC | D-ECE: 12.90 -> 11.18 (Sim10K); 9.54 -> 7.79 (KITTI); 9.40 -> 7.64 (Cityscapes); 15.42 -> 14.94 (COCO); 11.88 -> 6.02 (VOC) | Helped |
| Add MCCL train-time calibration to FCOS | FCOS · Out-of-domain shifts: Sim10K → Cityscapes, KITTI → Cityscapes, Cityscapes → Foggy Cityscapes, COCO → Cor-COCO, Cityscapes → BDD100K, VOC → Clipart, Watercolor, | D-ECE: 9.51 -> 6.60; 7.53 -> 6.43; 11.18 -> 8.97; 15.90 -> 14.45; 18.82 -> 16.12; 1.54 -> 1.06; 3.23 -> 2.33; 3.75 -> 2.84 | Helped |
| Add MCCL train-time calibration to Deformable DETR | Deformable DETR · In-domain and out-of-domain detection datasets | D-ECE: KITTI: 6.31 -> 3.87; lower values also reported across the other listed settings | Helped |
| Add localization calibration loss alongside multiclass confidence calibration loss (ablation) | MCCL / FCOS · COCO | D-ECE: 15.12 -> 14.94 | Helped |
| Add multiclass confidence calibration loss alongside localization calibration loss (ablation) | MCCL / FCOS · Sim10K → Cityscapes | D-ECE: 9.12 -> 6.60 | Helped |
| Add localization calibration loss alongside multiclass confidence calibration loss (ablation) | MCCL / FCOS · Sim10K → Cityscapes | D-ECE: 8.63 -> 6.60 | Helped |
| Remove both localization and multiclass confidence calibration losses (ablation) | MCCL / FCOS · COCO | D-ECE: 14.94 -> 15.26 | Hurt |
| Increase MC dropout passes from 3 to 15 (ablation) | MCCL / FCOS · COCO | D-ECE: 16.27 -> 16.12 | Helped |
| Increase MC dropout passes from 3 to 15 (ablation) | MCCL / FCOS · COCO-Corr. | D-ECE: 16.18 -> 16.17 | Mixed |
| Use MCCL rather than the baseline | FCOS · In-domain and out-of-domain datasets | AP@0.5 / mAP: In-domain AP@0.5: 87.45 -> 86.47 (Sim10K); 94.54 -> 93.76 (KITTI); 70.48 -> 70.22 (Cityscapes); 54.91 -> 54.85 (COCO); V | Mixed |
Closest papers by meaning
- Bridging Precision and Confidence: A Train-Time Loss for Calibrating Object Detection (CVPR 2023 · similarity 0.95)
- Towards Improving Calibration in Object Detection Under Domain Shift (NeurIPS 2022 · similarity 0.93)
- Cal-DETR: Calibrated Detection Transformer (NeurIPS 2023 · similarity 0.91)
- On Model Calibration for Long-Tailed Object Detection and Instance Segmentation (NeurIPS 2021 · similarity 0.88)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.87)
- Unsupervised Object Detection with Theoretical Guarantees (NeurIPS 2024 · similarity 0.86)
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.