Efficiently navigating to a specific kind of objects in an unknown environment is an important and challenging research topic in Embodied AI. Existing methods, such as end-to-end learning ones and modular ones still struggle at this task as they have poor efficiency, interpretability and/or generalization. This paper proposes an object-driven navigation strategy for robots based on active perception and semantic association. Specifically, this strategy enables robots to explore the environment and complete object-driven navigation tasks by endowing them with the ability to explore areas, infer associations, remember scenes, continue navigation, and so on. On this basis, this paper designs a semantic association model based on a graph convolution network using rich information hidden in the environment, including geometric and semantic information, to dynamically predict the association between objects, providing more accurate and effective prior knowledge for object-driven navigation. Experimental results prove that the proposed strategy can help robots better perceive and understand the environment and it is superior to similar strategies in terms of success rate and navigation efficiency.
Interactive segmentation is a technique for picking objects of interest in images according to users’ input interactions. Some recent works take the users’ interactive input to guide the deep neural network training, where the users’ click information is utilized as weak-supervised information. However, limited by the learning capability of the model, this structure does not accurately represent the user’s interaction intention. In this work, we propose a multi-click interactive segmentation solution for employing human intention to refine the segmentation results. We propose a coarse segmentation network to extract semantic information and generate rough results. Then, we designed a feature-aware attention module according to the symmetry of user intention and image semantic information. Finally, we establish a refinement module to combine the feature-aware results with coarse masks to generate precise intentional segmentation. Furthermore, the feature-aware module is trained as a plug-and-play tool, which can be embedded into most deep image segmentation models for exploiting users’ click information in the training process. We conduct experiments on five common datasets (SBD, GrabCut, DAVIS, Berkeley, MS COCO) and the results prove our attention module can improve the performance of image segmentation networks.
In recent years, deep learning-based Salient Object Detection (SOD) methods have made tremendous progress; however, their performance in complex scenarios has reached a bottleneck. In this paper, we propose a novel Progressive Difference Fusion Network (PDFNet) based on fine-grained feature fusion. First, to address the scale variability of salient objects, we introduce a Self-Guided Module (SGM) with dynamic receptive fields. Second, to tackle the shape variability of salient objects, we design a Feature Aggregation Module (FAM) incorporating cross convolutions and a feedback loop. Finally, to alleviate the issue of confusion between global and detail information during multi-scale feature fusion in existing models, we develop a Progressive Difference Fusion Unit (PDFU) to project multi-scale features into fine-grained nodes and enhance them through node interaction based on difference features. Additionally, we propose a Conditional Random Field Based on Patch (CRFbp), which focuses on handling discrete points, further improving the model’s performance. Extensive experiments demonstrate that our method achieves state-of-the-art (SOTA) performance on five benchmark datasets. Code is available at: https://github.com/pdfnet2025/PDFNet.git.