Multi-Modal Dynamic Graph Transformer for Visual Grounding

Sijia Chen, Baochun Li

Abstract

Visual Grounding (VG) aims to localize a target object in an image based on a natural language query. Traditional methods often treat this as a static ranking task, where performance is capped by the quality of initial region proposals. We present the Multi-Modal Dynamic Graph Transformer (M-DGT), a framework that reimagines visual grounding as a progressive search-and-refine process. By representing image regions as nodes in a dynamic graph and iteratively optimizing their spatial coordinates and semantic alignment, M-DGT achieves state-of-the-art results across multiple benchmarks while maintaining high inference efficiency.

Introduction

The core challenge in Visual Grounding (VG) is bridging the semantic gap between visual features and linguistic descriptions. Most existing "one-step" architectures follow an evaluate-and-rank paradigm: they either rely on pre-defined candidate regions or generate a single attention map that is refined only once. This approach is inherently limited; if the initial proposals are poor or the attention is slightly misaligned, the model lacks a mechanism to "fix" the bounding box iteratively.

To overcome this, we propose moving away from static classification toward a system that continuously optimizes the location and shape of regions. The Multi-Modal Dynamic Graph Transformer (M-DGT) treats the grounding task as a dynamic search, utilizing a graph-based representation to refine visual semantic alignment over multiple iterations.

The M-DGT Framework

Overview of the M-DGT architecture, showing the iterative flow from initial graph construction to spatial refinement and pruning.

Methodology

The M-DGT framework consists of three primary stages designed to transform a set of rough initial boxes into a precise target localization:

  1. Dynamic Graph Construction: The process begins with a small set of roughly initialized boxes (as few as 48) covering the image. These boxes serve as nodes in a graph, where edges represent spatial and semantic relationships between regions.
  2. Node Transformer: This module is the engine of the refinement process. It contains a Node-aware Cross-modal Encoder that fuses visual features (extracted via ROI Align) with linguistic features (from BERT) to evaluate the match between a box and the query. Subsequently, a Spatial-aware Decoder predicts 2D transformations—translation and scaling—to shift and resize the boxes toward the target.
  3. Graph Transformer: To maintain efficiency and resolve ambiguity, this module prunes the graph. It employs a "local-graph layout attention" mechanism to identify and remove redundant or irrelevant nodes as the iterations progress, focusing the model's computational resources on the most promising candidates.

Cross-Modal Alignment

Detailed view of the Node-aware Cross-modal Encoder, illustrating the fusion of visual and linguistic features.

A critical component of the Node Transformer is the cross-modal encoder, which ensures that the visual nodes are updated with relevant linguistic context. By calculating attention between text tokens and visual nodes, the model learns which regions are most relevant to specific parts of the query, facilitating more accurate spatial adjustments.

Experiments and Results

Performance Benchmarks

M-DGT was evaluated on four standard benchmarks: Flickr30k Entities, RefCOCO, RefCOCO+, and RefCOCOg. The model demonstrated significant improvements over existing methods:

  • Flickr30k Entities: Achieved a state-of-the-art accuracy of 79.97%, outperforming the previous best model (LCMCG) by over 3%.
  • RefCOCO Variants: Consistently outperformed leading methods like TransVG and VS-graph. Notably, M-DGT showed superior precision at high Intersection over Union (IoU) thresholds, indicating its ability to produce much tighter, more accurate bounding boxes.

Efficiency and Versatility

Despite its iterative nature, M-DGT remains computationally efficient. Using a ResNet101 backbone, it achieves an inference time of 108ms, making it competitive with one-step methods. Furthermore, M-DGT can function as a "refinement plugin" for other VG models. When applied to the outputs of models like SL-CCRF or LGRANs, it consistently boosted their accuracy by 1% to 1.5%.

Conclusion

The Multi-Modal Dynamic Graph Transformer shifts the paradigm of visual grounding from static ranking to iterative search. By leveraging dynamic graphs and pruning mechanisms, the model effectively resolves semantic ambiguity and overcomes the limitations of initial region proposals. Future work will explore deeper modeling of graph evolution across iterations and improving the handling of extremely complex linguistic relations.

Summary updated

Abstract

Visual grounding (VG) aims to align the correct regions of an image with a natural language query about that image. We found that existing VG methods are trapped by the single-stage grounding process that performs a sole evaluate-and-rank for meticulously prepared regions. Their performance depends on the density and quality of the candidate regions, and is capped by the inability to optimize the located regions continuously. To address these issues, we propose to remodel VG into a progressively optimized visual semantic alignment process. Our proposed multi-modal dynamic graph transformer (M-DGT) achieves this by building upon the dynamic graph structure with regions as nodes and their semantic relations as edges. Starting from a few randomly initialized regions, M-DGT is able to make sustainable adjustments (i.e., 2D spatial transformation and deletion) to the nodes and edges of the graph based on multi-modal information and the graph feature, thereby efficiently shrinking the graph to approach the ground truth regions. Experiments show that with an average of 48 boxes as initialization, the performance of M-DGT on the Flickr30k Entities and RefCOCO datasets outperforms existing state-of-the-art methods by a substantial margin, in terms of both accuracy and Intersect over Union (IOU) scores. Furthermore, introducing M-DGT to optimize the predicted regions of existing methods can further significantly improve their performance. The source codes are available at https://github.com/iQua/M-DGT.
BibTeX
@inproceedings{chen2022multi,
  author = {Sijia Chen and Baochun Li},
  title = {{Multi-Modal Dynamic Graph Transformer for Visual Grounding}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2022},
  doi = {10.1109/cvpr52688.2022.01509},
}
Figure