MEDPNet: Achieving High-Precision Adaptive Registration for Complex Die Castings

TL;DR

MEDPNet tackles the challenge of high-precision point cloud registration for complex die-cast parts by combining a coarse stage based on Efficient DCP (enhanced with Efficient Attention) with a refined stage MDR that fuses multi-scale ICP and NDT outputs through a learnable weighting scheme and an adaptive RMSE-driven mechanism for the final transform $T^*$. A die-casting–specific dataset, DieCastCloud, is introduced to evaluate performance under practical conditions, including significant surface features and noise. The results show MEDPNet achieving state-of-the-art accuracy and robustness on DieCastCloud, with efficient runtimes suitable for industrial deployment. Overall, the approach provides a scalable, noise-resistant, and adaptive registration framework for real-world die-casting applications, addressing both data scarcity and the need for precise alignment in manufacturing QA and molding processes.

Abstract

Due to their complex spatial structure and diverse geometric features, achieving high-precision and robust point cloud registration for complex Die Castings has been a significant challenge in the die-casting industry. Existing point cloud registration methods primarily optimize network models using well-established high-quality datasets, often neglecting practical application in real scenarios. To address this gap, this paper proposes a high-precision adaptive registration method called Multiscale Efficient Deep Closest Point (MEDPNet) and introduces a die-casting point cloud dataset, DieCastCloud, specifically designed to tackle the challenges of point cloud registration in the die-casting industry. The MEDPNet method performs coarse die-casting point cloud data registration using the Efficient-DCP method, followed by precision registration using the Multiscale feature fusion dual-channel registration (MDR) method. We enhance the modeling capability and computational efficiency of the model by replacing the attention mechanism of the Transformer in DCP with Efficient Attention and implementing a collaborative scale mechanism through the combination of serial and parallel blocks. Additionally, we propose the MDR method, which utilizes multilayer perceptrons (MLP), Normal Distributions Transform (NDT), and Iterative Closest Point (ICP) to achieve learnable adaptive fusion, enabling high-precision, scalable, and noise-resistant global point cloud registration. Our proposed method demonstrates excellent performance compared to state-of-the-art geometric and learning-based registration methods when applied to complex die-casting point cloud data.

Figures (5)

• Figure 1: Illustrates the point cloud registration process for a die-cast part. **"View 1" and "View 2" correspond to point clouds X and Y under different views.** Through an encoder, the point cloud data of the die-cast part are converted into data in feature space. Then, from the common parts of the unaligned point cloud pairs, the optimal transformation parameters R, t are calculated to obtain the best transformation matrix $T^*$, where R represents the rotation angles, t represents the translation components, and $T^*$ is the best transformation matrix.
• Figure 2: The architecture of MEDPNet. In the diagram, part (a) shows the structure of Efficient DCP, part (b) illustrates the composition of Efficient Attention, and part (c) outlines the framework of the MDR method, with a detailed exposition of its details. The MEDPNet method collects point cloud data of the same die-casting part under different postures through a robotic arm equipped with a 3D sensor, inputs the unaligned point cloud pairs into Efficient DCP for preliminary registration, and then refines the alignment through MDR to preserve essential feature information.
• Figure 3: Illustration of the architecture of efficient attention.Where the input feature map undergoes a transformation into three distinct components: Queries (Q), Keys (K), and Values (V). These components facilitate a self-attention schema by computing attention scores between Q and K, followed by a softmax normalization to acquire a probabilistic weight distribution. The weighted sum of these probabilities with V culminates in the output feature map, encapsulating a dynamic representation of salient features pivotal for subsequent layers of the network to process. This mechanism underpins the network's capacity to accentuate pertinent information within the feature space selectively.
• Figure 4: Adaptive parameter optimization. First, we input the formula to be optimized into an MLP and use backpropagation to optimize the weights. Next, through a self-updating filtering mechanism, we iterate to find the smallest registration error, iteratively updating the adaptive hyperparameter $\varepsilon$.
• Figure 5: Registration result visualization. We visualized the registration results of various methods under clean and noisy samples, ranking them in descending order according to the root mean square error.