DeepGCNs: Making GCNs Go as Deep as CNNs

Abstract

Convolutional Neural Networks (CNNs) have been very successful at solving a variety of computer vision tasks such as object classification and detection, semantic segmentation, activity understanding, to name just a few. One key enabling factor for their great performance has been the ability to train very deep networks. Despite their huge success in many tasks, CNNs do not work well with non-Euclidean data, which is prevalent in many real-world applications. Graph Convolutional Networks (GCNs) offer an alternative that allows for non-Eucledian data input to a neural network. While GCNs already achieve encouraging results, they are currently limited to architectures with a relatively small number of layers, primarily due to vanishing gradients during training. This work transfers concepts such as residual/dense connections and dilated convolutions from CNNs to GCNs in order to successfully train very deep GCNs. We show the benefit of using deep GCNs (with as many as 112 layers) experimentally across various datasets and tasks. Specifically, we achieve very promising performance in part segmentation and semantic segmentation on point clouds and in node classification of protein functions across biological protein-protein interaction (PPI) graphs. We believe that the insights in this work will open avenues for future research on GCNs and their application to further tasks not explored in this paper.

MAAS: Multi-modal Assignation for Active Speaker Detection

“…In computer vision, GCNs have been successfully applied to scene graph generation [22,31,38,52,56], 3D understanding [16,29,49,51], and action recognition in video [20,53,55]. In MAAS we desing a DeepGCN-like architecture [27,28,30], that addresses a special scenario, namely the multi-modal nature of audiovisual data. We rely on the well known EdgeConv operator [49], to model interactions between different modalities on graph nodes identified across multiple frames.…”

PointRGCN: Graph Convolution Networks for 3D Vehicles Detection Refinement

“…resentation fits the most with deep learning is not straightforward and remains an open problem [22,25,14]. Recent advances in graph convolution networks [14] suggest that graph representations could provide better features for point cloud processing. Such a representation already outperforms the state-of-the-art in many other computer vision tasks [32,19,28,34].…”

PU-GCN: Point Cloud Upsampling using Graph Convolutional Networks

“…To better represent locality, we leverage the power of graphs and specifically Graph Convolutional Networks (GCNs). GCNs are considered a versatile tool to process non-Euclidean data, and recent research on point cloud semantic and part segmentation shows their power in encoding local and global information [25,13,12]. In this paper, we use GCNs to design novel point cloud upsampling modules (refer to Figure 1), which are better equipped at encoding local information and learn to generate new point patches instead of merely replicating parts of the input.…”

“…To learn better hierarchical feature representation, graph pooling methods such as DIFFPool [29] and SAGPooling [11] are proposed. Recently, Li et al [13,12] introduced residual/skip connections and dilated convolutions to GCNs, and successfully trained high capacity GCN architectures over 100 layers in depth. Previous GCN works mainly investigate discriminative models for node classification or graph classification tasks.…”