Research map / Detection and segmentation / Open-set and few-shot detection
Domain Adaptation for Large-Vocabulary Object Detectors
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 KGD, which distills CLIP's implicit knowledge graphs (language via WordNet-enriched prompts, vision via dynamically updated object-feature graphs) into a Detic large-vocabulary detector to improve unsupervised domain adaptation pseudo-labels.
Problem
Large-vocabulary object detectors localize well but misclassify objects on downstream domains due to domain shift and vocabulary mismatch, producing noisy pseudo labels for unsupervised adaptation.
Approach
Extract language and vision knowledge graphs from CLIP (nodes = WordNet-enriched text prompts or visual class centroids, edges = feature distances), then use GCN-based reasoning over these graphs to calibrate the teacher detector's class probabilities for mean-teacher pseudo labeling.
Key results
- KGD reaches 53.6 AP50 on Cityscapes vs. 46.5 for the Detic baseline and 50.8 for the strongest CLIP-augmented baseline (SHOT+).
- KGD achieves the best AP50 on all 11 target datasets reported, e.g., 69.1 on Clipart1k vs. 61.0 for Detic.
- Ablation on Cityscapes: KGD-L alone gives 52.8 and KGD-V alone gives 52.7 AP50; combined 53.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply KGD adaptation | Detic · Cityscapes | AP50: 46.5 -> 53.6 | Helped |
| add KGD-L (ablation) | KGD adaptation · Cityscapes | AP50: 46.5 -> 52.8; +6.3 | Helped |
| add KGD-V (ablation) | KGD adaptation · Cityscapes | AP50: 46.5 -> 52.7; +6.2 | Helped |
| combine KGD-L and KGD-V (ablation) | KGD adaptation · Cityscapes | AP50: 52.8 -> 53.6; 52.7 -> 53.6 | Helped |
| use WordNet hierarchy instead of category names for LKG extraction (ablation) | KGD-L only · Cityscapes | AP50: 51.9 -> 52.8 | Helped |
| use WordNet hierarchy instead of Synset definitions for LKG extraction (ablation) | KGD-L only · Cityscapes | AP50: 52.0 -> 52.8 | Helped |
| use graph-based LKG encapsulation instead of feature-distance encapsulation (ablation) | KGD-L only · Cityscapes | AP50: 49.6 -> 52.8 | Helped |
| use dynamic VKG extraction instead of static VKG extraction (ablation) | KGD-V only · Cityscapes | AP50: 51.9 -> 52.7 | Helped |
| add smoothing to dynamic VKG extraction (ablation) | KGD-V only · Cityscapes | AP50: 52.2 -> 52.7 | Helped |
| vary pseudo-label threshold τ from 0.15 to 0.35 (ablation) | KGD · Cityscapes | AP50: 53.4; 53.2; 53.6; 53.9; 53.5 | No clear effect |
Surprising findings
- Plain mean teacher gains are modest (e.g., 46.5 to 49.1 on Cityscapes), and existing CLIP feature-distillation methods (VILD, RegionKD, OADP) gave less improvement than KGD.
Limitations
- Gains on common-object datasets (VOC, Objects365) are smaller, attributed to a smaller domain gap with LVD pretraining data.
- Relies on the teacher detector's proposals and CLIP crops; quality of pseudo labels depends on the base LVD.
Experimental setup
Datasets: LVIS ImageNet-21K Cityscapes Vistas BDD100K Pascal VOC Objects365 MIO-TCD BAAI VisDrone Clipart1k Watercolor2k
Benchmarks: Cityscapes Vistas BDD100K-weather BDD100K-time-of-day Pascal VOC Objects365 MIO-TCD BAAI VisDrone Clipart1k Watercolor2k Comic2k
Metrics: AP50
Baselines: Detic Mean Teacher SHOT SFOD HCL IRG-SFDA VILD RegionKD OADP
Builds on: Detic CLIP Mean Teacher WordNet GCN
Closest papers by meaning
- GCD: Advancing Vision-Language Models for Incremental Object Detection via Global Alignment and Correspondence Distillation (AAAI 2025 · similarity 0.87)
- Learning Background Prompts to Discover Implicit Knowledge for Open Vocabulary Object Detection (CVPR 2024 · similarity 0.86)
- CAKE: Category Aware Knowledge Extraction for Open-Vocabulary Object Detection (AAAI 2025 · similarity 0.86)
- How to Evaluate the Generalization of Detection? A Benchmark for Comprehensive Open-Vocabulary Detection (AAAI 2024 · similarity 0.86)
- Cyclic Contrastive Knowledge Transfer for Open-Vocabulary Object Detection (ICLR 2025 · similarity 0.85)
- DitHub: A Modular Framework for Incremental Open-Vocabulary Object Detection (NeurIPS 2025 · similarity 0.85)
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.