Research map / Vision-language and multimodal / Human and embodied modeling
ObjEmbed: Towards Universal Multimodal Object Embeddings
Shenghao Fu, Yukun Su, Fengyun Rao, Jing LYU, Xiaohua Xie, Wei-Shi Zheng
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 box IoU as the focal-loss label instead of binary labels (single token) (ablation) | ObjEmbed · COCO | AP: 37.1 -> 42.3 | Improved |
| decouple an object into separate class and IoU tokens (ablation) | ObjEmbed · COCO | AP: 42.3 -> 45.5 | Improved |
| place the class token before the IoU token (ablation) | ObjEmbed two-token design · COCO | AP: 45.1 -> 45.5 | No clear effect |
| add object instruction (Object i: <object><iou>) (ablation) | ObjEmbed · COCO | AP: 42.1 -> 45.5 | Improved |
| add task instructions on top of object instruction (ablation) | ObjEmbed · COCO | AP: 45.5 -> 47.1 | Improved |
| add image-level objective to object-level training (ablation) | ObjEmbed · local image retrieval (LIR*) | average score: 60.2 -> 62.9 | Improved |
| add object-level objective to image-level training (ablation) | ObjEmbed · global image retrieval (GIR) | average score: 81.2 -> 81.6 | Improved |
| use separate special tokens for global and local text embeddings (ablation) | ObjEmbed · COCO and RefCOCO | AP / accuracy: COCO +0.3 AP, RefCOCO +0.8 | Improved |
| supervise global image embedding with both short and long captions, with two global image tokens (ablation) | ObjEmbed · global image retrieval (GIR) | average score: 80.1 -> 81.6 | Improved |
| add box offset regression head on IoU embeddings (ablation) | ObjEmbed · COCO / RefCOCO | AP / accuracy: COCO 52.8 -> 52.5, RefCOCO 87.4 -> 86.2 | Hurt |
| increase number of proposals from 50 to 100 / 150 (ablation) | ObjEmbed · COCO | AP: 52.6 -> 53.0 (100), 51.9 (150) | No clear effect |
| mix ground-truth boxes into proposals (oracle) (ablation) | ObjEmbed-4B · COCO | AP: 53.0 -> 65.2 | Improved |
Closest papers by meaning
- VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks (ICLR 2025 · similarity 0.89)
- MMIU: Multimodal Multi-image Understanding for Evaluating Large Vision-Language Models (ICLR 2025 · similarity 0.88)
- CLIM: Contrastive Language-Image Mosaic for Region Representation (AAAI 2024 · similarity 0.87)
- Harnessing Frozen Unimodal Encoders for Flexible Multimodal Alignment (CVPR 2025 · similarity 0.87)
- Kernel-based Unsupervised Embedding Alignment for Enhanced Visual Representation in Vision-language Models (ICML 2025 · similarity 0.87)
- Composing Object Relations and Attributes for Image-Text Matching (CVPR 2024 · 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.
