Monocular 3D object detection refers to detecting 3D objects using a single camera. This approach offers low sensor costs, high resolution, and rich texture information, making it widely adopted. However, monocular sensors face challenges from environmental factors like occlusion and truncation, leading to reduced detection accuracy. Additionally, the lack of depth information poses significant challenges for predicting 3D positions. To address these issues, this paper presents a monocular 3D object detection method based on improvements to MonoCD, designed to enhance detection accuracy and robustness in complex environments. In order to effectively obtain and integrate depth information, this paper designs a multi-branch depth prediction with weight sharing module. Furthermore, an adaptive focus mechanism is proposed to emphasize target regions while minimizing interference from irrelevant areas. The experimental results demonstrate that MonoDFNet achieves significant improvements over existing methods, with AP3D gains of +4.09% (Easy), +2.78% (Moderate), and +1.63% (Hard), confirming its effectiveness in 3D object detection.
20 January 2020 in Neural Computing and Applications
Accurately predicting the remaining useful life of bearings is crucial for ensuring the reliability and safety of rotating machinery across various industrial scenarios. For instance, the bearings of air compressors in fuel cells, as key components, directly affect the reliability of the entire system. However, existing prediction methods still face three core challenges: insufficient extraction of degradation features, low network prediction accuracy, and inadequate local information acquisition capability of prediction networks. To address these issues, this paper proposes a novel deep residual shrinkage network-temporal convolutional network (DRSN-TCN) prediction framework integrated with weighted health indicator (HI) and a multi-scale attention mechanism, with its core innovations reflected in three aspects: first, a weighted feature fusion method is proposed to construct nonlinear HIs, which evaluates the contribution of each statistical feature to the degradation process based on three quantitative metrics—monotonicity, trendability, and prognosability, strengthens key degradation information and suppresses noise interference, and ultimately generates HIs with richer degradation information. Second, a DRSN-TCN integrated architecture is designed, where embedding the DRSN module into the TCN network not only expands network depth to enhance deep feature extraction capability but also mitigates the gradient explosion problem via residual learning and soft thresholding mechanisms. Third, a multi-scale attention mechanism is innovatively introduced to enable the model to simultaneously focus on the short-term fluctuations and long-term evolutionary trends of bearing degradation, thereby alleviating the local information loss issue of TCN caused by oversized dilated convolution kernels. Validation results based on two bearing durability test datasets show that, compared with current mainstream state-of-the-art methods, the proposed framework reduces the root mean square error by at least 31% and the mean absolute error by at least 24%, fully verifying its superior prediction accuracy and robustness.
Maximum a posteriori estimation (MAP) with Dirichlet prior has been shown to be effective in improving the parameter learning of Bayesian networks when the available data are insufficient. Given no extra domain knowledge, uniform prior is often considered for regularization. However, when the underlying parameter distribution is non-uniform or skewed, uniform prior does not work well, and a more informative prior is required. In reality, unless the domain experts are extremely unfamiliar with the network, they would be able to provide some reliable knowledge on the studied network. With that knowledge, we can automatically refine informative priors and select reasonable equivalent sample size (ESS). In this paper, considering the parameter constraints that are transformed from the domain knowledge, we propose a Constrained adjusted Maximum a Posteriori (CaMAP) estimation method, which is featured by two novel techniques. First, to draw an informative prior distribution (or prior shape), we present a novel sampling method that can construct the prior distribution from the constraints. Then, to find the optimal ESS (or prior strength), we derive constraints on the ESS from the parameter constraints and select the optimal ESS by cross-validation. Numerical experiments show that the proposed method is superior to other learning algorithms.