Manolis Savva

Example-based synthesis of 3D object arrangements

We present a method for synthesizing 3D object arrangements from examples. Given a few user-provided examples, our system can synthesize a diverse set of plausible new scenes by learning from a larger scene database. We rely on three novel contributions. First, we introduce a probabilistic model for scenes based on Bayesian networks and Gaussian mixtures that can be trained from a small number of input examples. Second, we develop a clustering algorithm that groups objects occurring in a database of scenes according to their local scene neighborhoods. These contextual categories allow the synthesis process to treat a wider variety of objects as interchangeable. Third, we train our probabilistic model on a mix of user-provided examples and relevant scenes retrieved from the database. This mixed model learning process can be controlled to introduce additional variety into the synthesized scenes. We evaluate our algorithm through qualitative results and a perceptual study in which participants judged synthesized scenes to be highly plausible, as compared to hand-created scenes.

Characterizing structural relationships in scenes using graph kernels

Modeling virtual environments is a time consuming and expensive task that is becoming increasingly popular for both professional and casual artists. The model density and complexity of the scenes representing these virtual environments is rising rapidly. This trend suggests that data-mining a 3D scene corpus could be a very powerful tool enabling more efficient scene design. In this paper, we show how to represent scenes as graphs that encode models and their semantic relationships. We then define a kernel between these relationship graphs that compares common virtual substructures in two graphs and captures the similarity between their corresponding scenes. We apply this framework to several scene modeling problems, such as finding similar scenes, relevance feedback, and context-based model search. We show that incorporating structural relationships allows our method to provide a more relevant set of results when compared against previous approaches to model context search.

Semantic Scene Completion from a Single Depth Image

This paper focuses on semantic scene completion, a task for producing a complete 3D voxel representation of volumetric occupancy and semantic labels for a scene from a single-view depth map observation. Previous work has considered scene completion and semantic labeling of depth maps separately. However, we observe that these two problems are tightly intertwined. To leverage the coupled nature of these two tasks, we introduce the semantic scene completion network (SSCNet), an end-to-end 3D convolutional network that takes a single depth image as input and simultaneously outputs occupancy and semantic labels for all voxels in the camera view frustum. Our network uses a dilation-based 3D context module to efficiently expand the receptive field and enable 3D context learning. To train our network, we construct SUNCG - a manually created largescale dataset of synthetic 3D scenes with dense volumetric annotations. Our experiments demonstrate that the joint model outperforms methods addressing each task in isolation and outperforms alternative approaches on the semantic scene completion task. The dataset and code is available at http://sscnet.cs.princeton.edu.

ReVision: automated classification, analysis and redesign of chart images

Poorly designed charts are prevalent in reports, magazines, books and on the Web. Most of these charts are only available as bitmap images; without access to the underlying data it is prohibitively difficult for viewers to create more effective visual representations. In response we present ReVision, a system that automatically redesigns visualizations to improve graphical perception. Given a bitmap image of a chart as input, ReVision applies computer vision and machine learning techniques to identify the chart type (e.g., pie chart, bar chart, scatterplot, etc.). It then extracts the graphical marks and infers the underlying data. Using a corpus of images drawn from the web, ReVision achieves image classification accuracy of 96% across ten chart categories. It also accurately extracts marks from 79% of bar charts and 62% of pie charts, and from these charts it successfully extracts data from 71% of bar charts and 64% of pie charts. ReVision then applies perceptually-based design principles to populate an interactive gallery of redesigned charts. With this interface, users can view alternative chart designs and retarget content to different visual styles.

ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes

A key requirement for leveraging supervised deep learning methods is the availability of large, labeled datasets. Unfortunately, in the context of RGB-D scene understanding, very little data is available – current datasets cover a small range of scene views and have limited semantic annotations. To address this issue, we introduce ScanNet, an RGB-D video dataset containing 2.5M views in 1513 scenes annotated with 3D camera poses, surface reconstructions, and semantic segmentations. To collect this data, we designed an easy-to-use and scalable RGB-D capture system that includes automated surface reconstruction and crowdsourced semantic annotation. We show that using this data helps achieve state-of-the-art performance on several 3D scene understanding tasks, including 3D object classification, semantic voxel labeling, and CAD model retrieval.

SceneGrok: inferring action maps in 3D environments

With modern computer graphics, we can generate enormous amounts of 3D scene data. It is now possible to capture high-quality 3D representations of large real-world environments. Large shape and scene databases, such as the Trimble 3D Warehouse, are publicly accessible and constantly growing. Unfortunately, while a great amount of 3D content exists, most of it is detached from the semantics and functionality of the objects it represents. In this paper, we present a method to establish a correlation between the geometry and the functionality of 3D environments. Using RGB-D sensors, we capture dense 3D reconstructions of real-world scenes, and observe and track people as they interact with the environment. With these observations, we train a classifier which can transfer interaction knowledge to unobserved 3D scenes. We predict a likelihood of a given action taking place over all locations in a 3D environment and refer to this representation as an action map over the scene. We demonstrate prediction of action maps in both 3D scans and virtual scenes. We evaluate our predictions against ground truth annotations by people, and present an approach for characterizing 3D scenes by functional similarity using action maps.

Learning Spatial Knowledge for Text to 3D Scene Generation

We address the grounding of natural language to concrete spatial constraints, and inference of implicit pragmatics in 3D environments. We apply our approach to the task of text-to-3D scene generation. We present a representation for common sense spatial knowledge and an approach to extract it from 3D scene data. In text-to3D scene generation, a user provides as input natural language text from which we extract explicit constraints on the objects that should appear in the scene. The main innovation of this work is to show how to augment these explicit constraints with learned spatial knowledge to infer missing objects and likely layouts for the objects in the scene. We demonstrate that spatial knowledge is useful for interpreting natural language and show examples of learned knowledge and generated 3D scenes.

Physically-Based Rendering for Indoor Scene Understanding Using Convolutional Neural Networks

Indoor scene understanding is central to applications such as robot navigation and human companion assistance. Over the last years, data-driven deep neural networks have outperformed many traditional approaches thanks to their representation learning capabilities. One of the bottlenecks in training for better representations is the amount of available per-pixel ground truth data that is required for core scene understanding tasks such as semantic segmentation, normal prediction, and object boundary detection. To address this problem, a number of works proposed using synthetic data. However, a systematic study of how such synthetic data is generated is missing. In this work, we introduce a large-scale synthetic dataset with 500K physically-based rendered images from 45K realistic 3D indoor scenes. We study the effects of rendering methods and scene lighting on training for three computer vision tasks: surface normal prediction, semantic segmentation, and object boundary detection. This study provides insights into the best practices for training with synthetic data (more realistic rendering is worth it) and shows that pretraining with our new synthetic dataset can improve results beyond the current state of the art on all three tasks.

Activity-centric scene synthesis for functional 3D scene modeling

We present a novel method to generate 3D scenes that allow the same activities as real environments captured through noisy and incomplete 3D scans. As robust object detection and instance retrieval from low-quality depth data is challenging, our algorithm aims to model semantically-correct rather than geometrically-accurate object arrangements. Our core contribution is a new scene synthesis technique which, conditioned on a coarse geometric scene representation, models functionally similar scenes using prior knowledge learned from a scene database. The key insight underlying our scene synthesis approach is that many real-world environments are structured to facilitate specific human activities, such as sleeping or eating. We represent scene functionalities through virtual agents that associate object arrangements with the activities for which they are typically used. When modeling a scene, we first identify the activities supported by a scanned environment. We then determine semantically-plausible arrangements of virtual objects -- retrieved from a shape database -- constrained by the observed scene geometry. For a given 3D scan, our algorithm produces a variety of synthesized scenes which support the activities of the captured real environments. In a perceptual evaluation study, we demonstrate that our results are judged to be visually appealing and functionally comparable to manually designed scenes.

PiGraphs: learning interaction snapshots from observations

We learn a probabilistic model connecting human poses and arrangements of object geometry from real-world observations of interactions collected with commodity RGB-D sensors. This model is encoded as a set of prototypical interaction graphs (PiGraphs), a human-centric representation capturing physical contact and visual attention linkages between 3D geometry and human body parts. We use this encoding of the joint probability distribution over pose and geometry during everyday interactions to generate interaction snapshots, which are static depictions of human poses and relevant objects during human-object interactions. We demonstrate that our model enables a novel human-centric understanding of 3D content and allows for jointly generating 3D scenes and interaction poses given terse high-level specifications, natural language, or reconstructed real-world scene constraints.