CVPR 2024 | Good at processing complex scenes and language expressions, Tsinghua & Bosch proposed a new instance segmentation network architecture MagNet

The AIxiv column is a column where this site publishes academic and technical content. In the past few years, the AIxiv column of this site has received more than 2,000 reports, covering top laboratories from major universities and companies around the world, effectively promoting academic exchanges and dissemination. If you have excellent work that you want to share, please feel free to contribute or contact us for reporting. Submission email: liyazhou@jiqizhixin.com; zhaoyunfeng@jiqizhixin.com.

Referring Image Segmentation (RIS) is a very challenging multi-modal task that requires the algorithm to be able to simultaneously understand fine-grained human language and visual image information, and segment the objects referred to by sentences in the image at the pixel level. Breakthroughs in RIS technology are expected to bring revolutionary changes in many fields such as human-computer interaction, image editing, and autonomous driving. It can greatly improve the efficiency and experience of human-machine collaboration. Although the current state-of-the-art RIS algorithms have made significant progress, they still face the problem of modality gap, that is, the distribution of image and text features are not completely aligned. This problem is particularly acute when dealing with complex referential language expressions and rare contexts.

^{Figure 1: Schematic diagram of the importance of fine-grained language-image alignment capabilities to RIS. The red mask is the prediction result of LAVT, one of the most advanced RIS algorithms currently, while the yellow dotted box is the correct annotation.}

Current RIS research mainly focuses on designing novel loss functions or introducing innovative network architectures/modules to enhance language-image distribution alignment . Despite significant progress, two fundamental issues remain that result in their inadequacy in fine-grained visual grounding:

1. These methods mainly rely on sentence-level language features for language-image alignment, resulting in their weak language-image alignment capabilities at the text level.

2. These methods often lack explicit supervision signals during the training process and cannot effectively teach the model to perform fine-grained alignment, causing them to perform poorly when processing complex referential languages.

##                                                                                                                                                                                                                                   In a recent CVPR 2024 work, a joint research team from the Department of Automation of Tsinghua University and Bosch Central Research Institute designed a new auxiliary task Mask Grounding. This task aims to explicitly teach the model to learn fine-grained correspondences between text and visual objects by randomly masking parts of text words and letting the algorithm learn to predict their true identities. In addition, they also proposed a novel cross-modal Alignment Module and a novel cross-modal alignment loss function (Cross-modal Alignment Loss) to further comprehensively reduce the gap between language and images. modal gap between. Based on these technologies, they designed a new instance segmentation network architecture Mask-grounded Network (MagNet).

Paper title: Mask Grounding for Referring Image Segmentation

Paper address: https://arxiv .org/abs/2312.12198

• On the RefCOCO, RefCOCO and G-Ref data sets, MagNet significantly surpassed all previous optimal algorithms, The core indicator of overall Interaction over Union (oIoU) increased significantly by 2.48 percentage points. The visualization results also confirm that MagNet has excellent performance in processing complex scenes and language expressions.

Method

##MagNet consists of 3 independent and complementary The modules are composed of Mask Grounding, Cross-modal Alignment Module and Cross-modal Alignment Loss.

1.Mask Grounding

                                                                                                                                                                                                           ^{Figure 3: Mask Grounding flow chart}

As shown in Figure 3, given the input image, the corresponding In the case of referential expressions and segmentation masks, the author randomly selects certain words in the sentence and replaces them with a special learnable mask token. The model is then trained to predict the actual identity of these replaced words. By successfully predicting the identity of the masked token, the model is able to understand which words in the text correspond to which parts of the image, learning fine-grained language-image alignment capabilities in the process. To perform this auxiliary task, the center coordinates of the mask region are first extracted and passed to a 2-layer MLP to encode the features of the segmentation mask. At the same time, a linear layer is used to map the language features to the same dimensions as the image features. Then, these features are jointly processed using the proposed mask token predictor, and the attention mechanism module is used for mask token prediction. Although Mask Grounding requires an additional forward pass through the language encoder to process the masked expressions, the overall computational cost is almost negligible because the language encoder is so small.

2.Cross-modal Alignment Module (CAM)

##                                                                                                                                                                                                               Figure 4: Cross-modal Alignment Module structure diagram

##As shown in Figure 4, in order to further improve the model performance, the author also proposed a cross-modal Alignment Module Modality Alignment Module (CAM), which enhances language-image alignment by injecting global context priors into image features before performing language-image fusion. CAM first generates K feature maps of different pyramid scales using pooling operations with different window sizes. Then, each feature map is passed through a 3-layer MLP to better extract global information and performs a cross-attention operation with another modality. Next, all output features are upsampled to the original feature map size by bilinear interpolation and concatenated in the channel dimension. Subsequently, a 2-layer MLP is used to reduce the number of concatenated feature channels back to the original dimensions. To prevent multimodal signals from overwhelming the original signal, a gated unit with Tanh nonlinearity is used to modulate the final output. Finally, this gated feature is added back to the input features and passed to the next stage of the image or language encoder. In the authors' implementation, CAM is added at the end of each stage of the image and speech encoders.

3.Cross-modal Alignment Loss (CAL)

##                                                                                                                                                                                                                                Figure 5: Cross-modal Alignment Loss formula

In order to supervise the model alignment of language and image features, the author proposes a novel cross-modal Alignment Loss formula State alignment loss function (CAL). Figure 5 shows the mathematical formula of this loss function. Unlike previous work, CAL considers both pixel-to-Pixel (P2P) and pixel-to-text (P2T) alignment. Accurate pixel-to-pixel alignment ensures that the model can segment and output segmentation masks with accurate shapes and boundaries, while accurate pixel-to-text alignment enables the model to correctly associate text descriptions with the image regions they match.

Experiment

In Table 1, the author evaluates MagNet using oIoU metric, and Performance comparison with existing state-of-the-art algorithms. The test data are RefCOCO, RefCOCO and G-Ref. In both single and multiple/additional dataset settings, MagNet's performance is all SOTA on these datasets.

                                                                                                                                                                                                                      ^{Table 1: Experimental results}

##Visualized results

                                                                                                                                                                                                                    Figure 6: MagNet visualization results

##In Figure 6, we can see that the visualization results of MagNet are also It stands out, outperforming the baseline LAVT in many difficult scenarios.

Summary

This article delves into the field of reference segmentation (RIS) challenges and current issues, especially the shortcomings in fine-grained language-image alignment. In response to these problems, researchers from Tsinghua University and Bosch Central Research Institute proposed a new method called MagNet, which comprehensively improves language by introducing the auxiliary task Mask Grounding, a cross-modal alignment module and a cross-modal alignment loss function. and the alignment effect between images. Experiments prove that MagNet achieves significantly better performance on the RefCOCO, RefCOCO and G-Ref data sets, surpassing the previous state-of-the-art algorithms and showing strong generalization capabilities. The visualization results also confirm the superiority of MagNet in processing complex scenes and language expressions. This research provides useful inspiration for the further development of the field of reference segmentation and is expected to promote greater breakthroughs in this field.

Team introduction

This paper comes from the Department of Automation, Tsinghua University (https:/ /www.au.tsinghua.edu.cn) and Bosch Central Research Institute (https://www.bosch.com/research/). One of the first authors of the paper, Zhuang Rongxian, is a doctoral student at Tsinghua University and is an intern at Bosch Academia Sinica; the project leader is Dr. Qiu Xuchong, a senior R&D scientist at Bosch Academia Sinica; the corresponding author is Professor Huang Gao from the Department of Automation, Tsinghua University.

The above is the detailed content of CVPR 2024 | Good at processing complex scenes and language expressions, Tsinghua & Bosch proposed a new instance segmentation network architecture MagNet. For more information, please follow other related articles on the PHP Chinese website!