Research map / Detection and segmentation / Open-set and few-shot detection
Region-Aware Pretraining for Open-Vocabulary Object Detection With Vision Transformers
Dahun Kim, Anelia Angelova, Weicheng Kuo
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 |
|---|---|---|---|
| Use CPE with focal contrastive loss instead of standard full-image PE and softmax cross-entropy (ablation) | RO-ViT pretraining · LVIS novel categories | AP_r: 21.4 -> 24.3; +2.9 AP_r | Improved |
| Add cropped positional embeddings (ablation) | RO-ViT pretraining · LVIS novel categories | AP_r: +2.4 AP_r | Improved |
| Replace softmax cross-entropy with focal contrastive loss (ablation) | RO-ViT pretraining · LVIS novel categories | AP_r: +0.5 AP_r | Improved |
| Remove positional embeddings (ablation) | RO-ViT pretraining · LVIS novel categories | AP_r: 21.4 -> 18.7 | Hurt |
| Replace learnable PE with sinusoidal PE (ablation) | RO-ViT pretraining · LVIS novel categories | AP_r: 21.4 -> 21.7 | No clear effect |
| Crop and resize the last feature map rather than positional embeddings (ablation) | RO-ViT pretraining · LVIS novel categories | AP_r: 21.4 -> 20.6 | Hurt |
| Add CPE and focal loss (ablation) | Frozen RO-ViT backbone · LVIS novel categories | AP_r: 9.7 -> 16.5; +6.8 AP_r | Improved |
| Set backbone finetuning LR ratio to 0.1 (ablation) | RO-ViT detector · LVIS novel categories | AP_r: 16.5 -> 24.3 | Improved |
| Use localization-quality objectness in detector scoring (ablation) | RO-ViT detector · LVIS novel categories | AP_r: +2.0 AP_r | Improved |
| Normalize classifier and mask output layers (ablation) | RO-ViT detector · LVIS novel categories | AP_r: +0.6 AP_r | Improved |
| Use CPE and focal loss instead of baseline pretraining (ablation) | RO-ViT ViT-B/16 · MS COCO retrieval | image-to-text Recall@1: 59.1 -> 62.2 | Improved |
| Increase backbone size from ViT-B/16 to ViT-L/16 (ablation) | RO-ViT with focal and CPE · MS COCO retrieval | image-to-text Recall@1: 62.2 -> 67.0 | Improved |
Closest papers by meaning
- Simple Open-Vocabulary Object Detection with Vision Transformers (ECCV 2022 · similarity 0.93)
- LLMs Meet VLMs: Boost Open Vocabulary Object Detection with Fine-grained Descriptors (ICLR 2024 · similarity 0.91)
- Cyclic Contrastive Knowledge Transfer for Open-Vocabulary Object Detection (ICLR 2025 · similarity 0.91)
- Exploring Region-Word Alignment in Built-in Detector for Open-Vocabulary Object Detection (CVPR 2024 · similarity 0.90)
- Language-conditioned Detection Transformer (CVPR 2024 · similarity 0.90)
- Generative Region-Language Pretraining for Open-Ended Object Detection (CVPR 2024 · 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.
