Research map / Detection and segmentation / CNN features
MI-DETR: An Object Detection Model with Multi-time Inquiries Mechanism
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 MI-DETR, a DETR-like decoder with parallel parameter-independent inquiry heads whose object queries are fused per layer, plus a U-like feature interaction module, improving COCO AP over DINO and Relation-DETR.
Problem
Cascaded decoder layers in DETR-like detectors restrict object queries to update in one direction, so they learn limited information from image features, which hurts detection of tiny, occluded, or background-confused objects on COCO.
Approach
Replace cascaded decoder layers with a parallel Multi-time Inquiries (MI) mechanism: M parameter-independent inquiry heads (self-attention, cross-attention, FFN) process object queries in parallel and their outputs are concatenated and linearly fused; a U-like Feature Interaction module feeds each encoder layer's features to the matching decoder layer as key/value.
Key results
- With ResNet-50 at 12 epochs, MI-DETR reaches 52.4 AP vs. 51.7 for Relation-DETR (+0.7) and 49.0 for DINO.
- With ResNet-50 at 24 epochs, MI-DETR reaches 52.7 AP vs. 52.1 for Relation-DETR (+0.6), the best reported in the paper.
- With Swin-L at 12 epochs, MI-DETR reaches 58.2 AP vs. 57.8 for Relation-DETR (+0.4).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MI to DINO (ablation) | DINO · COCO val2017, ResNet-50, 12 epochs | AP: 49.0 -> 50.2 | Helped |
| Add MI to Relation-DETR (ablation) | Relation-DETR · COCO val2017, ResNet-50, 12 epochs | AP: 51.7 -> 52.4 | Helped |
| Add MI only (ablation) | DINO · COCO val2017, ResNet-50, 12 epochs | AP: 49.0 -> 49.8 | Helped |
| Use Lite-MI instead of the baseline decoder (ablation) | DINO · COCO val2017, ResNet-50, 12 epochs | AP: 49.0 -> 49.6 | Helped |
| Add UFI only (ablation) | DINO · COCO val2017, ResNet-50, 12 epochs | AP: 49.0 -> 49.5 | Helped |
| Add MI and UFI (ablation) | DINO · COCO val2017, ResNet-50, 12 epochs | AP: 49.0 -> 50.1 | Helped |
| Add Lite-MI and UFI (ablation) | DINO · COCO val2017, ResNet-50, 12 epochs | AP: 49.0 -> 50.2 | Helped |
| Increase inquiry-head number from 1 to 4 (ablation) | MI-DETR · COCO val2017, ResNet-50, 12 epochs | AP: 49.5 -> 50.2 | Helped |
| Increase inquiry-head number from 4 to 5 (ablation) | MI-DETR · COCO val2017, ResNet-50, 12 epochs | AP: 50.2 -> 49.9 | Hurt |
| Use one inquiry head alone instead of fusing all four (ablation) | MI-DETR · COCO val2017, ResNet-50, 12 epochs | AP: 50.2 vs. 41.0, 42.2, 41.8, 40.4 | Hurt |
| Increase decoder layers from 6 to 24 (ablation) | DINO · COCO val2017, ResNet-50, 12 epochs | AP: 49.0 -> 47.3 | Hurt |
| Use 2 inquiry heads with 6 decoder layers instead of 12 decoder layers (ablation) | DINO · COCO val2017, ResNet-50, 12 epochs | AP: 49.0 -> 49.5 | Helped |
Surprising findings
- Simply stacking more decoder layers (DINO with 24 layers, 47.3 AP) degrades performance, while parallel inquiry heads on 6 layers improve it.
- Plugging MI into Relation-DETR at 12 epochs (52.4 AP) beats Relation-DETR trained for 24 epochs (52.1 AP), suggesting faster convergence.
Limitations
- Increasing inquiry heads beyond 4 degrades performance (IHN=5 gives 49.9 AP vs 50.2), so the head count needs tuning.
- The paper acknowledges that which specific patterns each inquiry head learns is hard to characterize; interpretation relies on visualization.
Experimental setup
Datasets: COCO train2017
Benchmarks: COCO val2017
Metrics: AP AP50 AP75 APS APM APL
Baselines: DINO Stable-DINO Rank-DETR Align-DETR DDQ-DETR DAC-DETR H-DETR Group-DETR MS-DETR Relation-DETR SQR-DETR
Builds on: DINO Relation-DETR DETR U-Net
Closest papers by meaning
- MS-DETR: Efficient DETR Training with Mixed Supervision (CVPR 2024 · similarity 0.89)
- Frozen-DETR: Enhancing DETR with Image Understanding from Frozen Foundation Models (NeurIPS 2024 · similarity 0.87)
- Order Matters: Competition-Guided Query Ordering for RNN-Based Object Detection (NeurIPS 2026 · similarity 0.87)
- DETRs Beat YOLOs on Real-time Object Detection (CVPR 2024 · similarity 0.86)
- Hybrid Proposal Refiner: Revisiting DETR Series from the Faster R-CNN Perspective (CVPR 2024 · similarity 0.86)
- CP-DETR: Concept Prompt Guide DETR Toward Stronger Universal Object Detection (AAAI 2025 · 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.