SeqTR: A Simple Yet Universal Network for Visual Grounding

Visual Grounding, also known as Referring Expression Comprehension and Phrase Grounding, aims to ground the specific region(s) within the image(s) based on the given expression text. This task simulates the common referential relationships between visual and linguistic modalities, enabling machines to develop human-like multimodal comprehension capabilities. Consequently, it has extensive applications in various domains. However, since 2021, visual grounding has witnessed significant advancements, with emerging new concepts such as grounded pre-training, grounding multimodal LLMs, generalized visual grounding, and giga-pixel grounding, which have brought numerous new challenges. In this survey, we first examine the developmental history of visual grounding and provide an overview of essential background knowledge, including fundamental concepts and evaluation metrics. We systematically track and summarize the advancements, and then meticulously define and organize the various settings to standardize future research and ensure a fair comparison. In the dataset section, we compile a comprehensive list of current relevant datasets, conduct a fair comparative analysis, and provide ultimate performance prediction to inspire the development of new standard benchmarks. Additionally, we delve into numerous applications and highlight several advanced topics. Finally, we outline the challenges confronting visual grounding and propose valuable directions for future research, which may serve as inspiration for subsequent researchers. By extracting common technical details, this survey encompasses the representative work in each subtopic over the past decade. To the best of our knowledge, this paper represents the most comprehensive overview currently available in the field of visual grounding. This survey is designed to be suitable for both beginners and experienced researchers, serving as an invaluable resource for understanding key concepts and tracking the latest research developments.

Text instance provides valuable information for the understanding and interpretation of natural scenes. The rich, precise high-level semantics embodied in the text could be beneficial for understanding the world around us, and empower a wide range of real-world applications. While most recent visual phrase grounding approaches focus on general objects, this paper explores extracting designated texts and predicting unambiguous scene text segmentation mask, i.e. scene text segmentation from natural language descriptions (referring expressions) like orange text on a little boy in black swinging a bat. The solution of this novel problem enables accurate segmentation of scene text instances from the complex background. In our proposed framework, a unified deep network jointly models visual and linguistic information by encoding both region-level and pixel-level visual features of natural scene images into spatial feature maps, and then decode them into saliency response map of text instances. To conduct quantitative evaluations, we establish a new scene text referring expression segmentation dataset: COCO-CharRef. Experimental results demonstrate the effectiveness of the proposed framework on the text instance segmentation task. By combining image-based visual features with language-based textual explanations, our framework outperforms baselines that are derived from state-of-the-art text localization and natural language object retrieval methods on COCO-CharRef dataset.

Visual grounding, a crucial vision-language task involving the understanding of the visual context based on the query expression, necessitates the model to capture the interactions between objects, as well as various spatial and attribute information. However, the annotation data of visual grounding task is limited due to its time-consuming and labor-intensive annotation process, resulting in the trained models being constrained from generalizing its capability to a broader domain. To address this challenge, we propose GroundVLP, a simple yet effective zero-shot method that harnesses visual grounding ability from the existing models trained from image-text pairs and pure object detection data, both of which are more conveniently obtainable and offer a broader domain compared to visual grounding annotation data. GroundVLP proposes a fusion mechanism that combines the heatmap from GradCAM and the object proposals of open-vocabulary detectors. We demonstrate that the proposed method significantly outperforms other zero-shot methods on RefCOCO/+/g datasets, surpassing prior zero-shot state-of-the-art by approximately 28% on the test split of RefCOCO and RefCOCO+. Furthermore, GroundVLP performs comparably to or even better than some non-VLP-based supervised models on the Flickr30k entities dataset. Our code is available at https://github.com/om-ai-lab/GroundVLP.

An object detector based on convolutional neural network (CNN) has been widely used in the field of computer vision because of its simplicity and efficiency. The average accuracy of CNN model detection results in the object detector is greatly affected by the loss function. The precision of the localization algorithm in the loss function is the main factor affecting the result. Based on the complete intersection over union (CIoU) loss function, an improved penalty function is proposed to improve the localization accuracy. Specifically, the algorithm more comprehensively considers matching bounding boxes between prediction with ground truth, using the proportional relationship of the aspect ratio from both bounding boxes. Under the same aspect ratio of the two bounding boxes, the influence factors of the prediction box on localization accuracy were considered. In this way, the function of the penalty function is strengthened, and localization accuracy of the network model improved. This loss function is called Improved CIoU (ICIoU). Experiments on the Udacity, PASCAL VOC, and MS COCO datasets have demonstrated the effectiveness of ICIoU in improving localization accuracy of network models by using the one-stage object detector YOLOv4. Compared with CIoU, the proposed ICIoU improved average precision (AP) by 0.57% and AP75 by 0.12% on Udacity, AP by 0.26% and AP75 by 1.28% on PASCAL VOC, and AP by 0.06% and AP75 by 0.65% on MS COCO.

Loss functions, such as the IoU Loss function and the GIoU (Generalized Intersection over Union) Loss function have been put forward to replace regression loss functions commonly used in regression loss calculation. GIoU Loss alleviates the vanishing gradient in the case of the non-overlapping, but it will completely degenerate into the IoU Loss function when bounding boxes overlap totally, which fails to achieve the optimization effect. To solve this problem, some improvements are proposed in this paper on the basis of the GIoU Loss function, taking into account the overlap rate of complete overlap of bounding boxes. In PASCAL VOC data, the experimental results demonstrate that the AP of NGIoU Loss function in the YOLOv4 model is 47.68%, 1.15% higher than that of the GIoU Loss function, and the highest map value is 86.79% in the YOLOv5 model.

Aerial image-based target detection has problems such as low accuracy in multiscale target detection situations, slow detection speed, missed targets and falsely detected targets. To solve this problem, this paper proposes a detection algorithm based on the improved You Only Look Once (YOLO)v3 network architecture from the perspective of model efficiency and applies it to multiscale image-based target detection. First, the K-means clustering algorithm is used to cluster an aerial dataset and optimize the anchor frame parameters of the network to improve the effectiveness of target detection. Second, the feature extraction method of the algorithm is improved, and a feature fusion method is used to establish a multiscale (large-, medium-, and small-scale) prediction layer, which mitigates the problem of small target information loss in deep networks and improves the detection accuracy of the algorithm. Finally, label regularization processing is performed on the predicted value, the generalized intersection over union (GIoU) is used as the bounding box regression loss function, and the focal loss function is integrated into the bounding box confidence loss function, which not only improves the target detection accuracy but also effectively reduces the false detection rate and missed target rate of the algorithm. An experimental comparison on the RSOD and NWPU VHR-10 aerial datasets shows that the detection effect of high-efficiency YOLO (HE-YOLO) is significantly improved compared with that of YOLOv3, and the average detection accuracies are increased by 8.92% and 7.79% on the two datasets, respectively. The algorithm not only shows better detection performance for multiscale targets but also reduces the missed target rate and false detection rate and has good robustness and generalizability.

While recent years have witnessed a dramatic upsurge of exploiting deep neural networks toward solving image denoising, existing methods mostly rely on simple noise assumptions, such as additive white Gaussian noise (AWGN), JPEG compression noise and camera sensor noise, and a general-purpose blind denoising method for real images remains unsolved. In this paper, we attempt to solve this problem from the perspective of network architecture design and training data synthesis. Specifically, for the network architecture design, we propose a swin-conv block to incorporate the local modeling ability of residual convolutional layer and non-local modeling ability of swin transformer block, and then plug it as the main building block into the widely-used image-to-image translation UNet architecture. For the training data synthesis, we design a practical noise degradation model which takes into consideration different kinds of noise (including Gaussian, Poisson, speckle, JPEG compression, and processed camera sensor noises) and resizing, and also involves a random shuffle strategy and a double degradation strategy. Extensive experiments on AGWN removal and real image denoising demonstrate that the new network architecture design achieves state-of-the-art performance and the new degradation model can help to significantly improve the practicability. We believe our work can provide useful insights into current denoising research. The source code is available at https://github.com/cszn/SCUNet.

The regression loss function is a key factor in the training and optimization process of object detection. The current mainstream regression loss functions are Ln norm loss, IOU loss and CIOU loss. This paper proposes the Scale-Sensitive IOU(SIOU) loss, a new loss function different from the above all, which could solve the issues that the current loss functions cannot distinguish the two bounding boxes in some special cases when the target area scales in one image vary greatly during training process, thereby leading to the improper regression loss calculation and the slowing down of the optimization. An area scale regulating factor Y is added on the basis of CIOU loss to adjust the loss values of the bounding boxes, which could distinguish all the boxes quantitatively in theory thus gets a faster converging speed and better optimization. Through analysis and simulation comparison among the several loss functions, the superiority of SIOU loss is verified. Furthermore, by incorporating SIOU loss into YOLO v4, Faster R-CNN and SSD on the two mainstream aerial remote sensing datasets, i.e., DIOR and NWPU VHR-10, the detection precisions improve by 10.2% than IOU loss and 2.8% than CIOU loss respectively.

It is challenging to accurately detect flexible objects with arbitrary orientation from monitoring images in power grid maintenance and inspection sites. This is because these images exhibit a significant imbalance between the foreground and background, which can lead to low detection accuracy when using a horizontal bounding box (HBB) as the detector in general object detection algorithms. Existing multi-oriented detection algorithms that use irregular polygons as the detector can improve accuracy to some extent, but their accuracy is limited due to boundary problems during the training process. This paper proposes a rotation-adaptive YOLOv5 (R_YOLOv5) with a rotated bounding box (RBB) to detect flexible objects with arbitrary orientation, effectively addressing the above issues and achieving high accuracy. Firstly, a long-side representation method is used to add the degree of freedom (DOF) for bounding boxes, enabling accurate detection of flexible objects with large spans, deformable shapes, and small foreground-to-background ratios. Furthermore, the further boundary problem induced by the proposed bounding box strategy is overcome by using classification discretization and symmetric function mapping methods. Finally, the loss function is optimized to ensure training convergence for the new bounding box. To meet various practical requirements, we propose four models with different scales based on YOLOv5, namely R_YOLOv5s, R_YOLOv5m, R_YOLOv5l, and R_YOLOv5x. Experimental results demonstrate that these four models achieve mean average precision (mAP) values of 0.712, 0.731, 0.736, and 0.745 on the DOTA-v1.5 dataset and 0.579, 0.629, 0.689, and 0.713 on our self-built FO dataset, exhibiting higher recognition accuracy and a stronger generalization ability. Among them, R_YOLOv5x achieves a mAP that is about 6.84% higher than ReDet on the DOTAv-1.5 dataset and at least 2% higher than the original YOLOv5 model on the FO dataset.

Embodied perception is essential for intelligent vehicles and robots in interactive environmental understanding. However, these advancements primarily focus on vision, with limited attention given to using 3D modeling sensors, restricting a comprehensive understanding of objects in response to prompts containing qualitative and quantitative queries. Recently, as a promising automotive sensor with affordable cost, 4D millimeter-wave radars provide denser point clouds than conventional radars and perceive both semantic and physical characteristics of objects, thereby enhancing the reliability of perception systems. To foster the development of natural language-driven context understanding in radar scenes for 3D visual grounding, we construct the first dataset, Talk2Radar, which bridges these two modalities for 3D Referring Expression Comprehension (REC). Talk2Radar contains 8,682 referring prompt samples with 20, 558 referred objects. Moreover, we propose a novel model, T-RadarNet, for 3D REC on point clouds, achieving State-Of-The-Art (SOTA) performance on the Talk2Radar dataset compared to counterparts. Deformable-FPN and Gated Graph Fusion are meticulously designed for efficient point cloud feature modeling and cross-modal fusion between radar and text features, respectively. Comprehensive experiments provide deep insights into radar-based 3D REC. We release our project at https://github.com/GuanRunwei/Talk2Radar.

We present a method for visually-grounded spoken term discovery. After training either a HuBERT or wav2vec2.0 model to associate spoken captions with natural images, we show that powerful word segmentation and clustering capability emerges within the model's self-attention heads. Our experiments reveal that this ability is not present to nearly the same extent in the base HuBERT and wav2vec2.0 models, suggesting that the visual grounding task is a crucial component of the word discovery capability we observe. We also evaluate our method on the Buckeye word segmentation and ZeroSpeech spoken term discovery tasks, where we outperform all currently published methods on several metrics. Code and model weights are available at https://github.com/jasonppy/word-discovery.

Existing benchmarks for visual question answering lack in visual grounding and complexity, particularly in evaluating spatial reasoning skills. We introduce FlowVQA, a novel benchmark aimed at assessing the capabilities of visual question-answering multimodal language models in reasoning with flowcharts as visual contexts. FlowVQA comprises 2,272 carefully generated and human-verified flowchart images from three distinct content sources, along with 22,413 diverse question-answer pairs, to test a spectrum of reasoning tasks, including information localization, decision-making, and logical progression. We conduct a thorough baseline evaluation on a suite of both open-source and proprietary multimodal language models using various strategies, followed by an analysis of directional bias. The results underscore the benchmark's potential as a vital tool for advancing the field of multimodal modeling, providing a focused and challenging environment for enhancing model performance in visual and logical reasoning tasks.

Accurate object detection requires correct classification and high-quality localization. Currently, most of the single shot detectors (SSDs) conduct simultaneous classification and regression using a fully convolutional network. Despite high efficiency, this structure has some inappropriate designs for accurate object detection. The first one is the mismatch of bounding box classification, where the classification results of the default bounding boxes are improperly treated as the results of the regressed bounding boxes during the inference. The second one is that only one-time regression is not good enough for high-quality object localization. To solve the problem of classification mismatch, we propose a novel reg-offset-cls (ROC) module including three hierarchical steps: the regression of the default bounding box, the prediction of new feature sampling locations, and the classification of the regressed bounding box with more accurate features. For high-quality localization, we stack two ROC modules together. The input of the second ROC module is the output of the first ROC module. In addition, we inject a feature enhanced (FE) module between two stacked ROC modules to extract more contextual information. The experiments on three different datasets (i.e., MS COCO, PASCAL VOC, and UAVDT) are performed to demonstrate the effectiveness and superiority of our method. Without any bells or whistles, our proposed method outperforms state-of-the-art one-stage methods at a real-time speed. The source code is available at https://github.com/JialeCao001/HSD.

We introduce a method to generate vectorial representations of visual classification tasks which can be used to reason about the nature of those tasks and their relations. Given a dataset with ground-truth labels and a loss function, we process images through a and compute an embedding based on estimates of the Fisher information matrix associated with the probe network parameters. This provides a fixed-dimensional embedding of the task that is independent of details such as the number of classes and requires no understanding of the class label semantics. We demonstrate that this embedding is capable of predicting task similarities that match our intuition about semantic and taxonomic relations between different visual tasks. We demonstrate the practical value of this framework for the meta-task of selecting a pre-trained feature extractor for a novel task. We present a simple meta-learning framework for learning a metric on embeddings that is capable of predicting which feature extractors will perform well on which task. Selecting a feature extractor with task embedding yields performance close to the best available feature extractor, with substantially less computational effort than exhaustively training and evaluating all available models.

We propose an end-to-end network for image generation from given structured-text that consists of the visual-relation layout module and stacking-GANs. Our visual-relation layout module uses relations among entities in the structured-text in two ways: comprehensive usage and individual usage. We comprehensively use all relations together to localize initial bounding-boxes (BBs) of all the entities. We use individual relation separately to predict from the initial BBs relation-units for all the relations. We then unify all the relation-units to produce the visual-relation layout, i.e., BBs for all the entities so that each of them uniquely corresponds to each entity while keeping its involved relations. Our visual-relation layout reflects the scene structure given in the input text. The stacking-GANs is the stack of three GANs conditioned on the visual-relation layout and the output of previous GAN, consistently capturing the scene structure. Our network realistically renders entities’ details while keeping the scene structure. Experimental results on two public datasets show the effectiveness of our method.