SegVG: Transferring Object Bounding Box to Segmentation for Visual Grounding

TL;DR

SegVG tackles visual grounding by fully exploiting the single box annotation through a bbox2seg paradigm, turning box-level supervision into pixel-level segmentation signals. It introduces a Multi-layer Multi-task Encoder-Decoder architecture and a Triple Alignment module (Tri-MHA) to iteratively ground the target via both regression and segmentation while harmonizing query, text, and vision features in a shared multimodal space. The approach yields state-of-the-art results on five standard datasets, with notable gains on challenging splits such as RefCOCO+ and RefCOCOg, and provides a segmentation-derived confidence score to improve reliability in practice. Overall, SegVG advances end-to-end, multi-task visual grounding by maximizing annotation utility and ensuring cross-modal alignment, enabling more accurate and dependable grounding in open-vocabulary settings.

Abstract

Different from Object Detection, Visual Grounding deals with detecting a bounding box for each text-image pair. This one box for each text-image data provides sparse supervision signals. Although previous works achieve impressive results, their passive utilization of annotation, i.e. the sole use of the box annotation as regression ground truth, results in a suboptimal performance. In this paper, we present SegVG, a novel method transfers the box-level annotation as Segmentation signals to provide an additional pixel-level supervision for Visual Grounding. Specifically, we propose the Multi-layer Multi-task Encoder-Decoder as the target grounding stage, where we learn a regression query and multiple segmentation queries to ground the target by regression and segmentation of the box in each decoding layer, respectively. This approach allows us to iteratively exploit the annotation as signals for both box-level regression and pixel-level segmentation. Moreover, as the backbones are typically initialized by pretrained parameters learned from unimodal tasks and the queries for both regression and segmentation are static learnable embeddings, a domain discrepancy remains among these three types of features, which impairs subsequent target grounding. To mitigate this discrepancy, we introduce the Triple Alignment module, where the query, text, and vision tokens are triangularly updated to share the same space by triple attention mechanism. Extensive experiments on five widely used datasets validate our state-of-the-art (SOTA) performance.

Figures (5)

• Figure 1: The comparison of visual grounding frameworks. The block with a dashed border indicates that the module may not necessarily exist. ( a) Previous baseline method consists of two backbones and additional transformer layers for target grounding, where a regression query is supervised to regress the box. Current SOTA methods further employ a text-to-visual module to align the visual features with text features. ( b) Our method incorporates segmentation queries, which utilizes the box annotation at the pixel-level to segment the target. Additionally, we propose the Triple Alignment module to eliminate the domain discrepancy of the query, text, and vision features.
• Figure 2: SegVG: The upper figure includes the vision and text backbone. Our proposed Triple Alignment module is iteratively inserted into intermediate layers to eliminate domain discrepancy. The lower figure shows our Multi-layer Multi-task Encoder-Decoder, which adopts a transformer encoder-decoder to update multimodal features and ground the target. In this architecture, we make the best of the box annotation as a segmentation ground truth and integrate an additional segmentation task into Visual Grounding. Additionally, the segmentation output serves as a Focal Loss factor, allowing adaptive emphasis on challenging cases for the regression loss. M = 6, R=6.
• Figure 3: IoU and Accuracy of different confidence scores on RefCOCOg-umd test set.
• Figure 4: Qualitative comparison between (c) (the first line) and (d) (the second line) of Table. \ref{['ablation']}. Red boxes are ground truth. Blue boxes are model predictions.
• Figure 5: Layer $i$ refers to the output of the $i$-th layer of the decoder. The blue boxes represent the models' predictions, while the red boxes denote the ground truth. In the segmentation mask shown in the second column from the right, red indicates high confidence for the foreground. Note that VLTVG does not provide segmentation output.