Research map / Detection and segmentation / CNN features

Order Matters: Competition-Guided Query Ordering for RNN-Based Object Detection

NeurIPS 2026 · Detection and segmentation › CNN features · approach: Region-based and query-based detectors · 0 citations (OpenAlex)

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 DETRNN, which orders DETR object queries by confidence-and-similarity (Competition-Guided Query Ordering) and refines them with a bidirectional RNN (Mamba by default) in place of decoder self-attention, reporting AP gains across several DETR-family detectors on COCO and CrowdHuman.

Problem

DETR-style detectors produce redundant, highly similar queries for the same object; decoder self-attention is parallel and order-agnostic, so it does not explicitly model competition among duplicates, which makes training unstable and predictions less decisive.

Approach

Build an explicit competition-aware query order from prior predictions (pivot-and-group by confidence and IoU/class-cosine similarity with density-adaptive thresholds), then replace decoder self-attention with a bidirectional RNN that refines queries along this order so that strong candidates suppress nearby competitors without discarding any queries.

Key results

• On COCO R50 1x with 900 queries, Def-DETRNN++ improves AP from 47.6 to 48.7 and DINO-RNN from 49.0 to 50.0.
• LoRA-DETRNN reaches 53.0 AP vs 52.5 for LoRA-DETR on COCO R50 1x while raising FPS from 16 to 19 at unchanged FLOPs.
• In a 1-decoder stress test, DETRNN raises Def-DETR++ AP from 40.6 to 46.3 and cuts the NMS gain from 2.4 to 0.3.

Surprising findings

• CGQO alone, without the recurrent module, gives no clear gain, and Bi-RNN alone gives only marginal gains; the improvement comes from their combination.
• Applying NMS@0.8 to the baselines yields only +0.1 to +0.2 AP on COCO and CrowdHuman, so the baselines already have few residual duplicates that NMS can remove.

Limitations

• Gains on full-recipe baselines are small (about +0.3 to +1.1 AP).
• CGQO grouping is partly sequential and costs about 0.48 ms at 900 queries, with most time in the grouping stage rather than the pairwise similarity.
• Results are mostly reported for ResNet-50 with 12-24 epochs; broader generalization is not established by the main tables.

Experimental setup

Datasets: MS-COCO 2017 CrowdHuman

Benchmarks: MS-COCO 2017 val2017 CrowdHuman val WiderPerson

Metrics: AP AP50 AP75 APS APM APL FPS FLOPs AR100 NMS AP gain

Baselines: Def-DETR++ DINO MS-DETR LoRA-DETR EASE-DETR Relation-DETR DDQ-DETR NMS@0.7/0.8/0.9

Builds on: DETR Mamba Deformable DETR DINO MS-DETR LoRA-DETR

Closest papers by meaning

• EASE-DETR: Easing the Competition among Object Queries (CVPR 2024 · similarity 0.89)
• MS-DETR: Efficient DETR Training with Mixed Supervision (CVPR 2024 · similarity 0.87)
• Hybrid Proposal Refiner: Revisiting DETR Series from the Faster R-CNN Perspective (CVPR 2024 · similarity 0.87)
• DetKDS: Knowledge Distillation Search for Object Detectors (ICML 2024 · similarity 0.87)
• MI-DETR: An Object Detection Model with Multi-time Inquiries Mechanism (CVPR 2025 · similarity 0.87)
• Salience DETR: Enhancing Detection Transformer with Hierarchical Salience Filtering Refinement (CVPR 2024 · similarity 0.86)