Research map / Vision-language and multimodal / Human and embodied modeling
TransVG++: End-to-End Visual Grounding With Language Conditioned Vision Transformer
Jiajun Deng, Zhengyuan Yang, Daqing Liu, Tianlang Chen, Wengang Zhou, Yanyong Zhang, Houqiang Li, Wanli Ouyang
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 |
|---|---|---|---|
| replace stand-alone fusion encoder layers with language adapters inside ViT-tiny (ablation) | TransVG++ · RefCOCOg val-g | accuracy (%): 67.24 -> 70.60 | Improved |
| replace stand-alone fusion with language adapter on ResNet-50 backbone (ablation) | TransVG · RefCOCOg val-g | accuracy (%): 66.35 -> 68.72 | Improved |
| replace ResNet-50 with ViT-tiny vision backbone (ablation) | TransVG · RefCOCOg val-g | accuracy (%): 66.35 -> 67.24 | Improved |
| use dilated ResNet-50 to match token count of ViT (ablation) | TransVG · RefCOCOg val-g | accuracy (%): 66.35 -> 65.96 | Hurt |
| use MHCA language adapter instead of average pooling (ablation) | TransVG++ · RefCOCOg val-g | accuracy (%): 69.49 -> 70.60 | Improved |
| use MHCA adapter instead of average pooling (ablation) | TransVG++ · ReferItGame val | accuracy (%): 72.01 -> 73.17 | Improved |
| place adapters evenly instead of in the first 4 layers (ablation) | TransVG++ · RefCOCOg val-g | accuracy (%): 68.26 -> 70.60 | Improved |
| place adapters evenly instead of the last 4 layers (ablation) | TransVG++ · RefCOCOg val-g | accuracy (%): 70.53 -> 70.60 | No clear effect |
| increase language conditioned layers from 2 to 4 (ablation) | TransVG++ · RefCOCOg val-g | accuracy (%): 70.13 -> 70.60 | Improved |
| increase language conditioned layers from 4 to 12 (ablation) | TransVG++ · RefCOCOg val-g | accuracy (%): 70.60 -> 70.30 | No clear effect |
| use learnable [REG] embedding instead of average pooled language tokens (ablation) | TransVG · ReferItGame val | accuracy (%): 69.96 -> 72.50 | Improved |
| add Transformers to both vision and language branches (ablation) | TransVG · ReferItGame test | accuracy (%): 64.24 -> 69.76 | Improved |
Closest papers by meaning
- TransVG: End-to-End Visual Grounding With Transformers (ICCV 2021 · similarity 0.95)
- SimVG: A Simple Framework for Visual Grounding with Decoupled Multi-modal Fusion (NeurIPS 2024 · similarity 0.93)
- UGround: Towards Unified Visual Grounding with Unrolled Transformers (ICML 2026 · similarity 0.89)
- GeM-VG: Towards Generalized Multi-image Visual Grounding with Multimodal Large Language Models (AAAI 2026 · similarity 0.89)
- VGent: Visual Grounding via Modular Design for Disentangling Reasoning and Prediction (CVPR 2026 · similarity 0.88)
- GroundVLP: Harnessing Zero-Shot Visual Grounding from Vision-Language Pre-training and Open-Vocabulary Object Detection (AAAI 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.
