In compositional zero-shot learning, the goal is to recognize unseen compositions (e.g. old dog) of observed visual primitives states (e.g. old, cute) and objects (e.g. car, dog) in the training set. This is challenging because the same state can for example alter the visual appearance of a dog drastically differently from a car. As a solution, we propose a novel graph formulation called Compositional Graph Embedding (CGE) that learns image features, compositional classifiers, and latent representations of visual primitives in an end-to-end manner. The key to our approach is exploiting the dependency between states, objects, and their compositions within a graph structure to enforce the relevant knowledge transfer from seen to unseen compositions. By learning a joint compatibility that encodes semantics between concepts, our model allows for generalization to unseen compositions without relying on an external knowledge base like WordNet. We show that in the challenging generalized compositional zero-shot setting our CGE significantly outperforms the state of the art on MIT-States and UT-Zappos. We also propose a new benchmark for this task based on the recent GQA dataset.
The distribution of a neural network's latent representations has been successfully used to detect Out-of-Distribution (OOD) data. Since OOD detection denotes a popular benchmark for epistemic uncertainty estimates, this raises the question of a deeper correlation. This work investigates whether the distribution of latent representations indeed contains information about the uncertainty associated with the predictions of a neural network. Prior work identifies epistemic uncertainty with the surprise, thus the negative log-likelihood, of observing a particular latent representation, which we verify empirically. Moreover, we demonstrate that the output-conditional distribution of hidden representations allows quantifying aleatoric uncertainty via the entropy of the predictive distribution. We analyze epistemic and aleatoric uncertainty inferred from the representations of different layers and conclude with the exciting finding that the hidden repesentations of a deterministic neural network indeed contain information about its uncertainty. We verify our findings on both classification and regression models.
Transferring the style from one image onto another is a popular and widely studied task in computer vision. Yet, learning-based style transfer in the 3D setting remains a largely unexplored problem. To our knowledge, we propose the first learning-based generative approach for style transfer between 3D objects. Our method allows to combine the content and style of a source and target 3D model to generate a novel shape that resembles in style the target while retaining the source content. The proposed framework can synthesize new 3D shapes both in the form of point clouds and meshes. Furthermore, we extend our technique to implicitly learn the underlying multimodal style distribution of the individual category domains. By sampling style codes from the learned distributions, we increase the variety of styles that our model can confer to a given reference object. Experimental results validate the effectiveness of the proposed 3D style transfer method on a number of benchmarks.
Domain generalization aims at training machine learning models to perform robustly across different and unseen domains. Several recent methods use multiple datasets to train models to extract domain-invariant features, hoping to generalize to unseen domains. Instead, first we explicitly train domain-dependant representations by using ad-hoc batch normalization layers to collect independent domain's statistics. Then, we propose to use these statistics to map domains in a shared latent space, where membership to a domain can be measured by means of a distance function. At test time, we project samples from an unknown domain into the same space and infer properties of their domain as a linear combination of the known ones. We apply the same mapping strategy at training and test time, learning both a latent representation and a powerful but lightweight ensemble model. We show a significant increase in classification accuracy over current state-of-the-art techniques on popular domain generalization benchmarks: PACS, Office-31 and Office-Caltech.
Estimating the 3D shape of an object from a single or multiple images has gained popularity thanks to the recent breakthroughs powered by deep learning. Most approaches regress the full object shape in a canonical pose, possibly extrapolating the occluded parts based on the learned priors. However, their viewpoint invariant technique often discards the unique structures visible from the input images. In contrast, this paper proposes to rely on viewpoint variant reconstructions by merging the visible information from the given views. Our approach is divided into three steps. Starting from the sparse views of the object, we first align them into a common coordinate system by estimating the relative pose between all the pairs. Then, inspired by the traditional voxel carving, we generate an occupancy grid of the object taken from the silhouette on the images and their relative poses. Finally, we refine the initial reconstruction to build a clean 3D model which preserves the details from each viewpoint. To validate the proposed method, we perform a comprehensive evaluation on the ShapeNet reference benchmark in terms of relative pose estimation and 3D shape reconstruction.
In this paper we propose a rotation-invariant deep network for point clouds analysis. Point-based deep networks are commonly designed to recognize roughly aligned 3D shapes based on point coordinates, but suffer from performance drops with shape rotations. Some geometric features, e.g., distances and angles of points as inputs of network, are rotation-invariant but lose positional information of points. In this work, we propose a novel deep network for point clouds by incorporating positional information of points as inputs while yielding rotation-invariance. The network is hierarchical and relies on two modules: a positional feature embedding block and a relational feature embedding block. Both modules and the whole network are proven to be rotation-invariant when processing point clouds as input. Experiments show state-of-the-art classification and segmentation performances on benchmark datasets, and ablation studies demonstrate effectiveness of the network design.
Panoptic segmentation has recently unified semantic and instance segmentation, previously addressed separately, thus taking a step further towards creating more comprehensive and efficient perception systems. In this paper, we present Panoster, a novel proposal-free panoptic segmentation method for point clouds. Unlike previous approaches relying on several steps to group pixels or points into objects, Panoster proposes a simplified framework incorporating a learning-based clustering solution to identify instances. At inference time, this acts as a class-agnostic semantic segmentation, allowing Panoster to be fast, while outperforming prior methods in terms of accuracy. Additionally, we showcase how our approach can be flexibly and effectively applied on diverse existing semantic architectures to deliver panoptic predictions.
Real-time scene reconstruction from depth data inevitably suffers from occlusion, thus leading to incomplete 3D models. Partial reconstructions, in turn, limit the performance of algorithms that leverage them for applications in the context of, e.g., augmented reality, robotic navigation, and 3D mapping. Most methods address this issue by predicting the missing geometry as an offline optimization, thus being incompatible with real-time applications. We propose a framework that ameliorates this issue by performing scene reconstruction and semantic scene completion jointly in an incremental and real-time manner, based on an input sequence of depth maps. Our framework relies on a novel neural architecture designed to process occupancy maps and leverages voxel states to accurately and efficiently fuse semantic completion with the 3D global model. We evaluate the proposed approach quantitatively and qualitatively, demonstrating that our method can obtain accurate 3D semantic scene completion in real-time.
3D Point clouds are a rich source of information that enjoy growing popularity in the vision community. However, due to the sparsity of their representation, learning models based on large point clouds is still a challenge. In this work, we introduce Graphite, a GRAPH-Induced feaTure Extraction pipeline, a simple yet powerful feature transform and keypoint detector. Graphite enables intensive down-sampling of point clouds with keypoint detection accompanied by a descriptor. We construct a generic graph-based learning scheme to describe point cloud regions and extract salient points. To this end, we take advantage of 6D pose information and metric learning to learn robust descriptions and keypoints across different scans. We Reformulate the 3D keypoint pipeline with graph neural networks which allow efficient processing of the point set while boosting its descriptive power which ultimately results in more accurate 3D registrations. We demonstrate our lightweight descriptor on common 3D descriptor matching and point cloud registration benchmarks and achieve comparable results with the state of the art. Describing 100 patches of a point cloud and detecting their keypoints takes only ~0.018 seconds with our proposed network.
This work proposes a RGB-D SLAM system specifically designed for structured environments and aimed at improved tracking and mapping accuracy by relying on geometric features that are extracted from the surrounding. Structured environments offer, in addition to points, also an abundance of geometrical features such as lines and planes, which we exploit to design both the tracking and mapping components of our SLAM system. For the tracking part, we explore geometric relationships between these features based on the assumption of a Manhattan World (MW). We propose a decoupling-refinement method based on points, lines, and planes, as well as the use of Manhattan relationships in an additional pose refinement module. For the mapping part, different levels of maps from sparse to dense are reconstructed at a low computational cost. We propose an instance-wise meshing strategy to build a dense map by meshing plane instances independently. The overall performance in terms of pose estimation and reconstruction is evaluated on public benchmarks and shows improved performance compared to state-of-the-art methods. We plan to publicly release the code of our SLAM framework.