Research map / Detection and segmentation / CNN features
Learning Dynamic Query Combinations for Transformer-based Object Detection and Segmentation
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 image-conditioned dynamic convex query combinations | DETR-based object detectors · MS COCO val | mAP: +0.8 to +2.3 mAP | Helped |
| Add image-conditioned dynamic convex query combinations | Mask2Former · CityScapes val | PQ: 62.1 -> 63.2; 66.1 -> 67.0 | Helped |
| Add image-conditioned dynamic convex query combinations | Mask2Former · MS COCO val | PQ: 51.9 -> 52.6; 55.1 -> 55.7 | Helped |
| Add image-conditioned dynamic convex query combinations | SeqFormer · YouTube-VIS-2019 val | mAP: 47.4 -> 49.0; 59.3 -> 61.2 | Helped |
| Constrain random query-combination coefficients to be convex rather than allow negative coefficients (ablation) | DAB-DETR, Deformable-DETR, and Mask2Former · MS COCO val | box or mask mAP: DAB-DETR r=2: 37.0$(\pm{0.10})$ -> 37.9$(\pm{0.10})$; Deformable-DETR r=2: 32.6$(\pm0.25)$ -> 35.0$(\pm{0.20})$; Mask2Fo | Helped |
| Use equal-weight averages instead of random convex query combinations (ablation) | DAB-DETR · MS COCO val | box mAP: 37.9$(\pm{0.10})$ -> 37.0 | Hurt |
| Generate queries directly from image features with an MLP instead of combining learned basic queries (ablation) | Deformable-DETR, ResNet-50 · MS COCO val | mAP: 46.2% -> 45.1% | Hurt |
| Remove the basic-query training loss by setting β=0 (ablation) | DQ-Deformable-DETR · MS COCO val | mAP: 47.0 -> 45.6 | Hurt |
| Use two unrelated query groups instead of learned modulated combinations (ablation) | Deformable-DETR · MS COCO val | mAP: DQ-Deformable-DETR 47.0; Deformable-DETR† 46.5 | Hurt |
| Use two unrelated query groups instead of learned modulated combinations (ablation) | Mask2Former · MS COCO val | mAP: DQ-Mask2Former 44.4; Mask2Former† 43.9 | Hurt |
| Reduce modulated queries from 300 to 100 (ablation) | DQ-Deformable-DETR and DQ-DAB-Deformable-DETR · MS COCO val | mAP | Helped |
| Use query ratio 4 rather than other tested ratios (ablation) | DQ-Deformable-DETR · MS COCO val | mAP | Helped |
Closest papers by meaning
- PaQ-DETR: Learning Pattern and Quality-Aware Dynamic Queries for Object Detection (CVPR 2026 · similarity 0.89)
- Anchor DETR: Query Design for Transformer-Based Detector (AAAI 2022 · similarity 0.89)
- Feature Aggregated Queries for Transformer-Based Video Object Detectors (CVPR 2023 · similarity 0.88)
- SAP-DETR: Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency (CVPR 2023 · similarity 0.87)
- DECO: Unleashing the Potential of ConvNets for Query-based Detection and Segmentation (ICLR 2025 · similarity 0.87)
- Rethinking Decoders for Transformer-based Semantic Segmentation: A Compression Perspective (NeurIPS 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.