The task of detecting three-dimensional objects using only RGB images presents a considerable challenge within the domain of computer vision. The core issue lies in accurately performing epipolar geometry matching between multiple views to obtain latent geometric priors. Existing methods establish correspondences along epipolar line features in voxel space through various layers of convolution. However, this step often occurs in the later stages of the network, which limits overall performance. To address this challenge, we introduce a novel framework, ImVoxelENet, that integrates a geometric epipolar constraint. We start from the back-projection of pixel-wise features and design an attention mechanism that captures the relationship between forward and backward features along the ray for multiple views. This approach enables the early establishment of geometric correspondences and structural connections between epipolar lines. Using ScanNetV2 as a benchmark, extensive comparative and ablation experiments demonstrate that our proposed network achieves a 1.1% improvement in mAP, highlighting its effectiveness in enhancing 3D object detection performance. Our code is available at https://github.com/xug-coder/ImVoxelENet.

With the development of deep learning techniques, learning implicit field for 3D shape reconstruction has become a heated topic, because implicit field can help networks learn a reasonable and sophisticated reconstruction model than explicit methods. However, there are still some challenges to be solved including lacking semantic information, local detail incompletions and so on. Thus, rather than recreating the entire model from a decoder directly, we first reconstruct the semantic components of a single model using an implicit filed structure based on semantic sections in our paper. Then we aggregate the reconstructed semantic parts together to get the final model. Finally, we test those results on the public 3D shape dataset PartNet and compare them to other cutting-edge single-view reconstruction approaches. It’s obvious that using a semantic part-based implicit field can learn more reasonable shape representations for reconstruction.