DC-DFFN: Densely Connected Deep Feature Fusion Network With Sign Agnostic Learning for Implicit Shape Representation

Reconstructing 3D surfaces from raw point cloud data is still a challenging and complex problem in computer vision and graphics. Recently emerged neural implicit representations model 3D surfaces implicitly in arbitrary resolution and diverse topologies. In this domain, most of the studies have so far used a single latent code-based variational auto-encoder (VAE) or auto-decoder (AD) architectures, or architectures similar to UNets. Due to the deep architectures of the existing approaches, gradients and/or input information can vanish while passing through the layers, which can cause suboptimal learning at training time and consequently low performance at test time. As a countermeasure, skip connections and feature fusion have been used in related application fields of convolutional neural networks. In this study, we embrace this idea and propose a novel densely connected deep feature fusion network, DC-DFFN, architecture for implicit shape representation. In the experimental results we show that DC-DFFN outperforms baseline approaches in terms visual reconstruction quality and quantitatively based on several measures. In addition, the proposed approach provides faster convergence during training compared to the baseline approaches. The DC-DFFN architecture has been implemented in PyTorch and is available as open source.