Research map / Detection and segmentation / Open-set and few-shot detection
Simple Multi-Dataset Detection
Xingyi Zhou, Vladlen Koltun, Philipp Krähenbühl
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 |
|---|---|---|---|
| Use dataset-specific heads and objectives with a shared backbone instead of simple dataset concatenation | Multi-dataset CascadeRCNN · Mean of COCO, Objects365, OpenImages | mAP: 33.2 -> 41.7 | Improved |
| Sample datasets uniformly rather than in proportion to dataset size (ablation) | Multi-dataset detector · Mean of COCO, Objects365, OpenImages | mAP: 33.2 -> 34.5 | Improved |
| Use class-aware sampling within Objects365 and OpenImages (ablation) | Multi-dataset detector · Objects365 | mAP: 16.5 -> 18.5 | Improved |
| Combine uniform dataset sampling with class-aware sampling (ablation) | Multi-dataset detector · Mean of COCO, Objects365, OpenImages | mAP: 38.5 -> 40.6 | Improved |
| Add a hierarchy-aware sigmoid cross-entropy loss for OpenImages (ablation) | Partitioned detector · OpenImages | mAP: +2.7 mAP | Improved |
| Extend the partitioned detector training schedule from 2× to 6× (ablation) | Partitioned detector · COCO | mAP: 41.8 -> 44.6 | Improved |
| Use an AP-based merge objective rather than GloVe word distance (ablation) | Unified label space · Mean of COCO, Objects365, OpenImages | mAP: 41.4{ $\pm$0.05} -> 41.9{ $\pm$0.02} | Improved |
| Learn the taxonomy from detector behavior rather than using the expert taxonomy (ablation) | Unified detector · Mean of COCO, Objects365, OpenImages | mAP: 41.6{ $\pm$0.04} -> 41.9{ $\pm$0.02} | Improved |
| Use output distortion instead of AP to learn the taxonomy (ablation) | Unified detector · Mean of COCO, Objects365, OpenImages | mAP: 41.9{ $\pm$0.02} -> 41.7{ $\pm$0.09} | Hurt |
| Vary the pruning threshold τ from 0.2 to 0.3 (ablation) | Unified label-space learning · COCO | mAP: 41.6 -> 41.8 | No clear effect |
| Retrain the detector on the unified taxonomy after weight merging (ablation) | Unified detector · COCO | mAP: 44.4 -> 45.4 | Improved |
| Train across four source datasets instead of using an ensemble of single-dataset models | Unified detector · Cross-dataset mean | mAP@0.5: 44.7 -> 47.3 | Improved |
Closest papers by meaning
- Detection Hub: Unifying Object Detection Datasets via Query Adaptation on Language Embedding (CVPR 2023 · similarity 0.91)
- ScaleDet: A Scalable Multi-Dataset Object Detector (CVPR 2023 · similarity 0.90)
- Object Detection with a Unified Label Space from Multiple Datasets (ECCV 2020 · similarity 0.89)
- "A Simple Approach and Benchmark for 21,000-Category Object Detection" (ECCV 2022 · similarity 0.88)
- Mixed Supervised Object Detection by Transferring Mask Prior and Semantic Similarity (NeurIPS 2021 · similarity 0.87)
- Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty (NeurIPS 2026 · 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.
