Research map / Detection and segmentation / Weakly supervised detection and segmentation
Sparse Semi-DETR: Sparse Learnable Queries for Semi-Supervised 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.
Sparse Semi-DETR improves DETR-based semi-supervised object detection by adding a Query Refinement Module that fuses multi-scale features into object queries and a Reliable Pseudo-Label Filtering Module that filters low-quality pseudo-labels for one-to-many assignment.
Problem
DETR-based semi-supervised object detection suffers from inaccurate pseudo-labels under one-to-one assignment, overlapping duplicate predictions under one-to-many assignment, and poor detection of small or occluded objects.
Approach
Refine object queries by aggregating high-resolution weakly-augmented features with attention and similarity fusion (Query Refinement Module), and filter unreliable pseudo-boxes using augmented ground-truth groups with top-k selection (Reliable Pseudo-Label Filtering), built on Semi-DETR with a DINO detector.
Key results
- On COCO-Partial 10% labels with ResNet-50, Sparse Semi-DETR reaches 44.3 mAP vs 43.5 for Semi-DETR (+0.8).
- On COCO-Full, Sparse Semi-DETR reaches 51.3 mAP, up from 49.2 supervised-stage result, surpassing Semi-DETR's 50.4.
- On Pascal VOC, Sparse Semi-DETR achieves 86.30 AP50 and 65.51 AP50:95 vs Semi-DETR's 86.10 and 65.20.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Sparse Semi-DETR instead of Semi-DETR | DINO · COCO-Partial 10% labels | mAP: 0.80 mAP | Helped |
| add Sparse Semi-DETR instead of Semi-DETR | DINO · COCO-Full with unlabel2017 | mAP: +2.1 | Helped |
| add Sparse Semi-DETR instead of Semi-DETR | DINO · Pascal VOC12 | AP50:95: 5.91 | Helped |
| add Query Refinement (ablation) | Sparse Semi-DETR · COCO-Partial 10% labels | mAP: 43.5 -> 43.8 | Helped |
| add Pseudo-Label Filtering (ablation) | Sparse Semi-DETR · COCO-Partial 10% labels | mAP: 43.5 -> 43.9 | Helped |
| apply Query Refinement to both student and teacher (ablation) | Sparse Semi-DETR · COCO-Partial 10% labels | mAP: 43.5 -> 42.8 | Hurt |
| add MLP for Ft1 and Fs1 (ablation) | Query Refinement · COCO-Partial 10% labels | mAP: 44.3 -> 43.7 | Hurt |
| replace Query Refinement with single-view queries (ablation) | Sparse Semi-DETR · COCO-Partial 10% labels | mAP: 44.3 -> 43.0 | Hurt |
| replace Query Refinement with cross-view queries (ablation) | Sparse Semi-DETR · COCO-Partial 10% labels | mAP: 44.3 -> 43.5 | Hurt |
| replace attention block with simple concatenation (ablation) | Query Refinement · COCO-Partial 10% labels | mAP: 44.3 -> 43.4 | Hurt |
| replace attention block with cosine similarity (ablation) | Query Refinement · COCO-Partial 10% labels | mAP: 44.3 -> 43.7 | Hurt |
| set augmented-ground-truth count m to 8 instead of 6 (ablation) | Pseudo-Label Filtering · COCO-Partial 10% labels | mAP: 44.3 -> 44.2 | Hurt |
Surprising findings
- Applying the query refinement strategy to strongly augmented teacher features degrades performance (teacher+student QR gives 42.8 mAP vs 44.3 for student-only QR).
Limitations
- Duplicate predictions remain when one-to-many assignment is applied in both stages; NMS is needed for the 120k-iteration variant.
- Gains over Semi-DETR are modest (about 0.4 to 0.8 mAP on COCO-Partial).
Experimental setup
Datasets: MS-COCO train2017 MS-COCO unlabel2017 Pascal VOC2007 Pascal VOC2012
Benchmarks: MS-COCO COCO-Partial (1%, 5%, 10%) MS-COCO COCO-Full Pascal VOC (VOC12 protocol)
Metrics: AP50:95 AP50 AP75 APS APM APL
Baselines: Semi-DETR Omni-DETR Dense Teacher PseCo DSL Soft Teacher Unbiased Teacher v2 Instant-Teaching Humble Teacher STAC
Builds on: Semi-DETR DINO DN-DETR Hybrid-DETR
Closest papers by meaning
- Semi-DETR: Semi-Supervised Object Detection With Detection Transformers (CVPR 2023 · similarity 0.95)
- Unbiased Teacher v2: Semi-Supervised Object Detection for Anchor-Free and Anchor-Based Detectors (CVPR 2022 · similarity 0.90)
- Points As Queries: Weakly Semi-Supervised Object Detection by Points (CVPR 2021 · similarity 0.89)
- Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity (ICLR 2022 · similarity 0.89)
- MS-DETR: Efficient DETR Training with Mixed Supervision (CVPR 2024 · similarity 0.87)
- Multi-clue Consistency Learning to Bridge Gaps Between General and Oriented Object in Semi-supervised Detection (AAAI 2025 · similarity 0.87)
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.