Research map / Detection and segmentation / Open-set and few-shot detection
Open-vocabulary Object Detection via Vision and Language Knowledge Distillation
Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, Yin Cui
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add image-embedding distillation to text-embedding classification (ablation) | ViLD on LVIS v1 · LVIS v1 novel categories | mask AP_r: 10.1 -> 16.1 | Improved |
| Use CLIP joint image-text embeddings instead of GloVe text embeddings (ablation) | ViLD-text on LVIS v1 · LVIS v1 novel categories | mask AP_r: 3.0 -> 10.1 | Improved |
| Ensemble ViLD-text and ViLD-image predictions using separate heads (ablation) | ViLD on LVIS v1 · LVIS v1 novel categories | mask AP_r: 16.1 -> 16.6 | Improved |
| Use ALIGN instead of CLIP as the teacher (ablation) | ViLD-ensemble on LVIS v1 · LVIS v1 novel categories | mask AP_r: 16.6 -> 26.3 | Improved |
| Add proposal objectness to CLIP confidence scores (ablation) | CLIP on cropped regions · LVIS v1 novel categories | mask AP_r: 13.0 -> 18.9 | Improved |
| Use a single prompt instead of ensembling multiple templates and synonyms (ablation) | ViLD on LVIS v1 · LVIS v1 novel categories | mask AP_r: 15.7 -> 16.1 | Hurt |
| Add L1 image-embedding distillation to the model without distillation (ablation) | ViLD on LVIS v1 · LVIS v1 novel categories | mask AP_r: 10.4 -> 12.9; 10.4 -> 14.0; 10.4 -> 16.3; 10.4 -> 17.3 | Improved |
| Use L1 rather than L2 for image-embedding distillation (ablation) | ViLD on LVIS v1 · LVIS v1 novel categories | mask AP_r: L1: 12.9, 14.0, 16.3, 17.3; L2: 13.7, 12.4, 13.4 | Improved |
| Increase the L1 distillation weight (ablation) | ViLD on LVIS v1 · LVIS v1 novel categories | mask AP_r: 12.9 -> 14.0 -> 16.3 -> 17.3 | Improved |
| Use image-embedding distillation alongside text classification (ablation) | ViLD on COCO · COCO novel categories | AP: 5.9 -> 27.6 | Improved |
| Transfer the LVIS-trained detector without finetuning rather than finetune (ablation) | ViLD on PASCAL VOC 2007 test · PASCAL VOC 2007 test | box AP_50: 78.9 vs 72.2 | Hurt |
| Transfer the LVIS-trained detector without finetuning rather than finetune (ablation) | ViLD on COCO validation · COCO validation | box AP: 39.1 vs 36.6 | Hurt |
Closest papers by meaning
- DitHub: A Modular Framework for Incremental Open-Vocabulary Object Detection (NeurIPS 2025 · similarity 0.92)
- Cyclic Contrastive Knowledge Transfer for Open-Vocabulary Object Detection (ICLR 2025 · similarity 0.91)
- Generative Region-Language Pretraining for Open-Ended Object Detection (CVPR 2024 · similarity 0.91)
- GCD: Advancing Vision-Language Models for Incremental Object Detection via Global Alignment and Correspondence Distillation (AAAI 2025 · similarity 0.91)
- LLMs Meet VLMs: Boost Open Vocabulary Object Detection with Fine-grained Descriptors (ICLR 2024 · similarity 0.91)
- CAKE: Category Aware Knowledge Extraction for Open-Vocabulary Object Detection (AAAI 2025 · similarity 0.90)
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.
