Research map / Detection and segmentation / CNN features
EASE-DETR: Easing the Competition among Object Queries
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EASE-DETR, a plug-in that encodes pairwise leading/trailing query relations (ranking plus IoU) into decoder self-attention and cross-attention decay weights to ease query competition and improve DETR-family detectors on COCO.
Problem
DETR-style detectors must make non-duplicate predictions via a hard competition among many redundant object queries around each object, where the leading query must dramatically widen its score gap over trailing queries across decoder layers, slowing training and limiting accuracy.
Approach
At each intermediate decoder layer, compute a binary leading/trailing relation between query pairs (score ranking), multiply it with IoU, and learn MLP-based decay weights that attenuate suppression on the leading query in self-attention and reduce feature absorption by trailing queries in cross-attention, adding almost no compute.
Key results
- EASE-Deformable++ with ResNet-50 improves 12-epoch AP from 47.0 to 48.3 (+1.3).
- EASE-DETR (DINO-based, with DAC-DETR one-to-many) reaches 50.8 AP on ResNet-50 at 12 epochs and 57.8 AP with Swin-L.
- On PETR multi-view 3D detection, EASE improves NDS by 1.6 (42.6 to 44.2) and mAP by 0.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the EASE module | Deformable++ · COCO 2017 val, 12 epochs, R50 | AP: 47.0 -> 48.3 (+1.3) | Helped |
| Add the EASE module | H-DETR · COCO 2017 val, 12 epochs, Swin-T | AP: 50.6 -> 51.7 (+1.1) | Helped |
| Add the EASE module | DINO · COCO 2017 val, 12 epochs, R50 | AP: 49.0 -> 49.7 (+0.7) | Helped |
| Use EASE attention with the one-to-many strategy | DINO · COCO 2017 val, 12 epochs, R50 | AP: 49.0 -> 50.8 (+1.8) | Helped |
| Use EASE attention with the one-to-many strategy | DINO · COCO 2017 val, 24 epochs, R50 | AP: 50.4 -> 51.6 (+1.2) | Helped |
| Use EASE attention with the one-to-many strategy | DINO · COCO 2017 val, 12 epochs, Swin-L | AP: 56.8 -> 57.8 (+1.0) | Helped |
| Add MSelf-attention (ablation) | Deformable++ · COCO 2017 val, 12 epochs, R50 | AP: 47.0 -> 48.1; +1.1 AP | Helped |
| Add MCross-attention after MSelf-attention (ablation) | Deformable++ · COCO 2017 val, 12 epochs, R50 | AP: 48.1 -> 48.3; +0.2 AP | Helped |
| Modulate self-attention using Rrank only (ablation) | Deformable++ · COCO 2017 val, 12 epochs, R50 | AP: 47.0 -> 47.8 (+0.8) | Helped |
| Modulate self-attention using Rrank multiplied by Rspt (ablation) | Deformable++ · COCO 2017 val, 12 epochs, R50 | AP: 47.0 -> 48.1 (+1.1) | Helped |
| Add MSelf-attention to the two-stage Deformable DETR configuration (ablation) | Two-stage Deformable DETR with DP0, MQS, and LFT · COCO | AP: 47.0 -> 48.1 (+1.1) | Helped |
| Add MQS to the MSelf-attention configuration (ablation) | Two-stage Deformable DETR with DP0 and LFT · COCO | AP: 47.4 -> 47.7 (+0.3) | Helped |
Surprising findings
- Adding one-to-many supervision on top of EASE-DETR is needed for the best numbers; the plain Deformable++ gain from cross-attention modulation is only +0.2 AP.
Limitations
- Gains on strong baselines are modest (often under 1 AP in some settings).
- Best SOTA numbers rely on combining with one-to-many supervision from DAC-DETR rather than EASE alone.
Experimental setup
Datasets: COCO 2017 nuScenes
Benchmarks: COCO 2017 val nuScenes
Metrics: AP AP50 AP75 APS APM APL NDS mAP GFLOPs
Baselines: Deformable++ DETR H-DETR DINO Stable-DINO DAC-DETR Group-DETR Align-DETR PETR DAB-DETR
Builds on: DETR Deformable DETR DINO DAB-DETR PETR
Closest papers by meaning
- DAC-DETR: Divide the Attention Layers and Conquer (NeurIPS 2023 · similarity 0.89)
- Order Matters: Competition-Guided Query Ordering for RNN-Based Object Detection (NeurIPS 2026 · similarity 0.89)
- Salience DETR: Enhancing Detection Transformer with Hierarchical Salience Filtering Refinement (CVPR 2024 · similarity 0.87)
- MI-DETR: An Object Detection Model with Multi-time Inquiries Mechanism (CVPR 2025 · similarity 0.86)
- MS-DETR: Efficient DETR Training with Mixed Supervision (CVPR 2024 · similarity 0.86)
- AQ-DETR: Low-Bit Quantized Detection Transformer with Auxiliary Queries (AAAI 2024 · similarity 0.83)
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.