Go to ICLR 2025 Conference homepageCTNet: A CNN-Transformer Hybrid Network for 6D Object Pose Estimation
Recent advances in 6D pose estimation primarily rely on CNNs, but they struggle to grasp long-range dependencies and the global context, which are essential for precise pose determination. Although deeper or expanded networks are commonly used to tackle this, they lead to significant computational burdens without fully addressing these constraints. To overcome these challenges, we present CTNet, a hybrid network that fuses the strengths of CNN and Transformer, aiming for accurate 6D pose estimation from a solitary RGB-D image. CTNet employs Transformer to capture elusive long-range dependencies and the global context, while lightweight CNNs adeptly extract detailed local features. This complementary approach offers a comprehensive feature representation, eliminating the necessity for excessively deep networks. To further bolster the CNNs' efficiency, we introduce the Hierarchical Feature Extractor (HFE), which enhances the C2f and ELAN modules for optimal feature extraction. Additionally, we integrate a CNN-based PointNet module, designed to extract vital spatial data from the point cloud. The Transformer element captures global contextual insights, which are then seamlessly integrated with the local and spatial features extracted by the CNNs to ensure precise 6D pose estimation. Experiments demonstrate that CTNet achieves high accuracy with nearly half the FLOPs of current methods on the LineMOD and YCB-Video datasets. Furthermore, the HFE is highly adaptable, showing excellent transferability across other 6D pose estimation architectures.