Research map / Detection and segmentation / Open-set and few-shot detection
CQ-DINO: Mitigating Gradient Dilution via Category Queries for Vast Vocabulary Object Detection
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 CQ-DINO, a GroundingDINO-based detector that replaces classification heads with learnable category queries selected per image by cross-attention, to tackle gradient dilution in vast vocabulary detection.
Problem
Classification-head detectors suffer positive and hard-negative gradient dilution when the category vocabulary grows to over 10^4 classes (e.g., V3Det with 13,204 categories).
Approach
Reformulate classification as contrastive matching between object queries and learnable category queries (initialized from CLIP text embeddings); an image-guided cross-attention selector keeps only the top-C' categories per image, rebalancing gradients and implicitly mining hard negatives, with hierarchical tree pooling for V3Det and self-attention for COCO.
Key results
- On V3Det with Swin-B-22k, CQ-DINO reaches 52.3 AP, outperforming Prova (50.3 AP) and DINO (43.4 AP).
- On COCO val2017 with Swin-L, CQ-DINO reaches 58.5 AP in 24 epochs, comparable to DETR variants such as DINO (58.0) and Rank-DETR (58.2).
- Image-guided query selection raises V3Det AP from 47.3 to 51.1 and FPS from 0.7 to 10.8 on the Swin-B-22k setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CQ-DINO instead of Prova | V3Det with Swin-B · V3Det val | AP: 44.5 -> 46.3 | Helped |
| Use CQ-DINO instead of Prova | V3Det with Swin-B-22k · V3Det val | AP: 50.3 -> 52.3 | Helped |
| Use CQ-DINO instead of Prova | V3Det with Swin-L · V3Det val | AP: 50.9 -> 53.0 | Helped |
| Use CQ-DINO instead of Rank-DETR | COCO val2017 with Swin-L · COCO val2017 | AP: 58.2 -> 58.5 | Helped |
| Add image-guided query selection to the baseline (ablation) | CQ-DINO on V3Det · V3Det, Swin-B-22k ablation | AP: 47.3 -> 51.1 | Helped |
| Add hierarchical tree construction (ablation) | CQ-DINO with image-guided selection on V3Det · V3Det, Swin-B-22k ablation | AP: +1.2% AP | Helped |
| Use self-attention instead of no category correlation encoding (ablation) | CQ-DINO with image-guided selection on V3Det · V3Det, Swin-B-22k ablation | AP: +0.2% AP | Helped |
| Add self-attention category relation learning (ablation) | CQ-DINO on COCO · COCO | AP: 58.3 -> 58.5 | Helped |
| Replace fixed tree weight (0.5) with adaptive weighting (ablation) | CQ-DINO on V3Det · V3Det tree-weight ablation | AP: 51.9 -> 52.3 | Helped |
| Tune DINO Focal Loss to α=0.50, γ=2 (ablation) | DINO with Swin-B-22k · V3Det | AP: 43.4 -> 47.4 | Helped |
| Tune DINO Focal Loss to α=0.50, γ=2 (ablation) | DINO with Swin-B · V3Det | AP: 42.0 -> 38.7 | Hurt |
| Tune DINO Focal Loss to α=0.50, γ=2 (ablation) | DINO with Swin-L · V3Det | AP: 48.5 -> 50.1 | Helped |
Surprising findings
- Focal Loss hyperparameters tuned on Swin-B-22k (alpha=0.5) degrade Swin-B performance by 3.3 AP, so the optimal setting is architecture-dependent.
- Generation-based open-ended detectors score very low on V3Det (GenerateU 0.4 AP, ChatRex 1.3 AP) zero-shot.
Limitations
- Performance depends on the recall of the category query selection; AR^C is 83.3% on V3Det.
- Two-stage training paradigm may yield suboptimal coordination compared with end-to-end training.
- Small-object performance on low-recall categories is weak (14.1 AP).
Experimental setup
Datasets: V3Det COCO CLIP text embeddings for initialization
Benchmarks: V3Det COCO val2017
Metrics: AP AP50 AP75 AR^C FPS
Baselines: DINO Deformable DETR Prova Faster R-CNN Cascade R-CNN GenerateU ChatRex Rank-DETR
Builds on: GroundingDINO DINO Query2Label Asymmetric Loss
Closest papers by meaning
- CAKE: Category Aware Knowledge Extraction for Open-Vocabulary Object Detection (AAAI 2025 · similarity 0.88)
- SHiNe: Semantic Hierarchy Nexus for Open-vocabulary Object Detection (CVPR 2024 · similarity 0.87)
- Cyclic Contrastive Knowledge Transfer for Open-Vocabulary Object Detection (ICLR 2025 · similarity 0.86)
- DitHub: A Modular Framework for Incremental Open-Vocabulary Object Detection (NeurIPS 2025 · similarity 0.86)
- Open-Vocabulary Object Detection via Language Hierarchy (NeurIPS 2024 · similarity 0.86)
- Comprehensive Multi-Modal Prototypes Are Simple and Effective Classifiers for Vast-Vocabulary 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.