Search

In recent years, general object detection on nature images leverages convolutional neural networks (CNNs) and vision transformers (ViT) to achieve great progress and high accuracy. However, unlike nature scenarios, remote sensing systems usually deploy a large number of edge devices. Therefore, this condition encourages detectors to have lower parameters than high-complexity neural networks. To this end, we propose an effective and lightweight detection framework of hybrid network, which enhances representation learning to balance efficiency and accuracy of model. Specifically, to compensate low precision caused by lightweight neural networks, we design a boundary-aware context (BAC) module and a frequency self-attention refinement (FSAR) module to improve detector performance in a hybrid structure. The BAC module enhances the local features of the object by fusing the spatial context information with the original image, which not only improves the model accuracy but also effectively solves the problem of multiscale objects. To alleviate the interference of complex background, the FSAR module adopts an adaptive technology to filter out redundant information at different frequencies to improve overall detection performance. The comprehensive experiments on remote sensing datasets, i.e., NWPU VHR-10, LEVIR, and RSOD, indicate that the proposed method achieves state-of-the-art performance and balances between model size and accuracy.

The existing RGB-D saliency object detection (SOD) methods mostly explore the complementary information between depth features and RGB features. However, these methods ignore the bi-directional complementarity between RGB and depth features. From this view, we propose a joint attention learning network (JALNet) to learn the cross-modal mutual complementary effect between the RGB images and depth maps. Specifically, two joint attention learning networks are designed, namely, a cross-modal joint attention fusion module (JAFM) and a joint attention enhance module (JAEM), respectively. The JAFM learns cross-modal complementary information from the RGB and depth features, which can strengthen the interaction of information and complementarity of useful information. At the same time, we utilise the JAEM to enlarge receptive field information to highlight salient objects. We conducted comprehensive experiments on four public datasets, which proved that the performance of our proposed JALNet outperforms 16 state-of-the-art (SOTA) RGB-D SOD methods.

Monocular depth estimation is a fascinating and challenging problem in virtual vision. However, the training of networks based on deep learning largely depends on the training data. This paper proposes a depth prediction method based on the depth cue: atmospheric light scattering, which can effectively predict the depth in different atmospheric light scenarios. But the assumption of global atmospheric light constancy can produce the unavoidable error. Especially for complex scenes, the complex reflected light of the scene leads to uneven distribution of atmospheric light. This paper proposes a new local atmospheric light estimation method, which can simulate the real distribution of atmospheric light scattering in the air more effectively. And the experiments found that the two models are complementary. In order to fuse the intrinsic real information of the two models, this paper adopts the fusion strategy based on the Bayesian model, and edge-preserving filtering is used to preserve the detailed information.

Cross-modal complementary information from RGB images and depth maps has brought new vitality and progress into salient object detection (SOD). Most existing RGB-D SOD methods generate saliency results by exploring the additional information of depth features, or multi-modality information from RGB and depth data. However, these methods ignore the mutual role between RGB and depth features. Different from them, we reconsider the status of the two modalities and propose an Encoder Steered Multi-modality Feature Guidance Network (EFGNet) to explore the mutual role of the two modalities for RGB-D salient object detection. To this end, a bi-directional framework based on encoder steered strategy is employed to extract and enhance the unimodal features with the aid of another modal data via the circular interactions on the encoder strategy. Specifically, a Multi-modality Feature Guided Module (M2FGM) is proposed to achieve the exploration of multi-modality mutual role, which uses depth features to guide and enhance RGB features (Depth→RGB), in turn, uses enhanced RGB features to guide and enhance depth features (RGB→Depth). Furthermore, we also design a Deep Feature-guided Decoder (DGD), which constructs a guidance block by embedding deep decoder features. Comprehensive experiments on six public datasets demonstrate that the proposed EFGNet outperforms state-of-the-art (SOTA) RGB-D SOD methods. Particularly, our method can obtain the percentage gain of 11.1% in term of MAE score on large-scale STERE dataset. Furthermore, the EFGNet runs 160 FPS with 160M model size when testing a 256×256 image on a single NVIDIA 2080Ti GPU. [ABSTRACT FROM AUTHOR]