Single-Stage Object 6D Pose Estimation Algorithm Using Multi-Scale Dynamic Optimization Strategy

In robotic motion control, accurate target pose estimation is essential for precise grasping. While deep learning techniques have advanced detection accuracy, they often suffer from slow detection speeds and require large amounts of training data. To overcome these limitations, this paper enhances a single-stage pose estimation algorithm by focusing on optimizing the detector's structure, particularly the feature fusion between the Neck and Backbone components. The study introduces a multi-scale dynamic optimization approach that integrates Dynamic Snake Convolution and Dysample technology. Dynamic Snake Convolution adapts the shape and positioning of convolution kernels to more effectively capture target features, while Dysample is an efficient, lightweight dynamic upsampler that learns upsampling through adaptive sampling. These advancements improve both the accuracy and speed of the detector without significantly increasing computational demands. To evaluate the proposed method, we performed a series of ablation and comparison experiments.