Research map / Detection and segmentation / Open-set and few-shot detection
GCD: Advancing Vision-Language Models for Incremental Object Detection via Global Alignment and Correspondence Distillation
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 GCD, a distillation framework for incremental object detection with vision-language detectors (Grounding DINO) that combines global alignment with pseudo labels and semantic-correspondence-based response and topology distillation on COCO.
Problem
Vision-language detectors forget old categories in incremental object detection because local alignment and disjoint embedding spaces cause semantic structure collapse; traditional KD does not transfer well to VLDs due to text-dependent encoding and decoding inconsistencies.
Approach
Align old and new objects-texts globally in one embedding space (old knowledge supervised by teacher pseudo labels), use a shared query plus chunked text tokens so teacher and student produce corresponding predictions, distill responses (CRD) and prototype topology (CTD).
Key results
- On COCO 70+10 two-phase setting, GCD reaches 46.7 AP versus 42.9 for TALIR and 40.9 for SSDGR (vision baseline).
- On COCO 40+40 two-phase setting, GCD reaches 45.7 AP, above SSDGR's 43.0.
- In multi-phase 40+10x4, GCD achieves 40.2 AP vs 36.8 for SSDGR; ablation shows FPP drops from 8.3 (local) to 1.9 (full GCD).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GCD instead of the strongest reported baseline | Grounding Dino · COCO 2017 70+10 | AP: 42.9 -> 46.7 | Helped |
| Use GCD instead of the strongest reported baseline | Grounding Dino · COCO 2017 40+40 | AP: 43.0 -> 45.7 | Helped |
| Use GCD instead of SSDGR | Grounding Dino · COCO 2017 40+10×4 | AP: 36.8 -> 40.2 | Helped |
| Use GCD instead of SSDGR | Grounding Dino · COCO 2017 40+20×2 | AP: 41.1 -> 44.0 | Helped |
| Add pseudo-labels to full-text global alignment (ablation) | Grounding Dino · COCO 2017 70+10 | AP: 6.3 -> 45.3 | Helped |
| Add CRD to global alignment with pseudo-labels (ablation) | Grounding Dino · COCO 2017 70+10 | AP: 45.3 -> 46.0 | Helped |
| Add CTD to global alignment + CRD (ablation) | Grounding Dino · COCO 2017 70+10 | FPP: 2.9 -> 1.9 | Helped |
| Replace CRD with raw response distillation (ablation) | Global alignment with pseudo-labels · COCO 2017 70+10 | AP: 45.3 -> 42.9 | Hurt |
Surprising findings
- Naively applying response distillation on top of the global alignment baseline decreases performance (42.9 vs 45.3 AP in the ablation), motivating the correspondence mechanism.
- Traditional KD-based ERD reproduced on Grounding DINO performs poorly in multi-phase settings (7.2 AP in 40+10x4).
Limitations
- Evaluated only on COCO 2017 with class-split incremental settings.
- Relies on pseudo labels and a global confidence threshold, which can still introduce noise.
Experimental setup
Datasets: COCO 2017
Benchmarks: COCO 2017
Metrics: AP AP50 AP75 APS APM APL FPP
Baselines: CL-DETR DMD SSDGR TALIR LWF ERD
Builds on: Grounding DINO ERD CL-DETR TALIR
Closest papers by meaning
- Cyclic Contrastive Knowledge Transfer for Open-Vocabulary Object Detection (ICLR 2025 · similarity 0.89)
- DitHub: A Modular Framework for Incremental Open-Vocabulary Object Detection (NeurIPS 2025 · similarity 0.89)
- Re-Aligning Language to Visual Objects with an Agentic Workflow (ICLR 2025 · similarity 0.88)
- Open-Det: An Efficient Learning Framework for Open-Ended Detection (ICML 2025 · similarity 0.88)
- Exploring Region-Word Alignment in Built-in Detector for Open-Vocabulary Object Detection (CVPR 2024 · similarity 0.88)
- LLMs Meet VLMs: Boost Open Vocabulary Object Detection with Fine-grained Descriptors (ICLR 2024 · similarity 0.88)
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.