InteriorNet: Mega-scale Multi-sensor Photo-realistic Indoor Scenes Dataset (1809.00716v1)

Abstract: Datasets have gained an enormous amount of popularity in the computer vision community, from training and evaluation of Deep Learning-based methods to benchmarking Simultaneous Localization and Mapping (SLAM). Without a doubt, synthetic imagery bears a vast potential due to scalability in terms of amounts of data obtainable without tedious manual ground truth annotations or measurements. Here, we present a dataset with the aim of providing a higher degree of photo-realism, larger scale, more variability as well as serving a wider range of purposes compared to existing datasets. Our dataset leverages the availability of millions of professional interior designs and millions of production-level furniture and object assets -- all coming with fine geometric details and high-resolution texture. We render high-resolution and high frame-rate video sequences following realistic trajectories while supporting various camera types as well as providing inertial measurements. Together with the release of the dataset, we will make executable program of our interactive simulator software as well as our renderer available at https://interiornetdataset.github.io. To showcase the usability and uniqueness of our dataset, we show benchmarking results of both sparse and dense SLAM algorithms.

• The paper presents a mega-scale dataset featuring 22 million interior layouts and over 1 million CAD models to enhance indoor scene understanding.
• It employs advanced photorealistic rendering and simulated dynamic environments to generate realistic visual and sensor data.
• The dataset supports rigorous SLAM evaluation with realistic camera trajectories, integrated IMU readings, and event camera outputs.

An In-Depth Analysis of the InteriorNet Dataset for Computer Vision Applications

The paper "InteriorNet: Mega-scale Multi-sensor Photo-realistic Indoor Scenes Dataset" introduces a substantial contribution to the field of computer vision, particularly in the domains of Simultaneous Localization and Mapping (SLAM) and data-driven learning for spatial perception. In addressing significant scalability challenges faced by existing datasets, this work develops a comprehensive synthetic dataset comprising a multitude of interior layouts and furniture models, designed to provide a high degree of photo-realism and extensive diversity for indoor scene understanding.

Central to this dataset is its foundation on a vast compilation of 1,042,632 CAD models obtained from leading manufacturers, with high-resolution textures and accurate real-world dimensions. These models are categorized into 158 main classes and are compatible with the NYU40 semantic categories. The interior layouts, totaling approximately 22 million, are crafted by over a thousand professional designers, underscoring their applicability in real-world decoration scenarios.

Dataset and Rendering Innovations

InteriorNet's design offers several notable advancements:

• Enormous Scale and Diversity: The dataset provides around 1 million furniture CAD models and includes 22 million meticulously designed interior layouts. This scale ensures a representative spread of common domestic environments.
• Photorealistic Rendering: Utilizing the ExaRenderer, the dataset showcases a rendering framework capable of producing high-resolution image sequences at high frame rates. The path tracing-based rendering facilitates realistic global illumination, supporting diverse lighting models and dynamic scene adjustments.
• Simulated Dynamic Environments: Through integration with a physics engine, the dataset includes dynamic rearrangement of movable objects, replicating day-to-day environmental changes. Additionally, variations in lighting conditions across different scenarios are supported to mimic real-world variability.

Trajectory Generation and Simulation Tools

The paper further details the generation of camera trajectories, which are crucial for SLAM and robotic navigation tasks. Three types of trajectories are constructed, with an overlay of a learned style model to impart realistic camera jitter, capturing the nuanced motion seen in practical scenarios. These synthetic trajectories are further augmented with inertial (IMU) readings and event camera outputs, thereby enhancing the dataset's utility for testing advanced sensor fusion algorithms.

The accompanying simulator, ViSim, allows for user-driven trajectory configuration, supporting a wide range of camera models and facilitating seamless integration with existing SLAM evaluation pipelines.

SLAM Evaluation and Practical Applications

To affirm the dataset's efficacy, the authors present an empirical evaluation using ORBSLAM2.0 and ElasticFusion. A series of benchmarking experiments demonstrated the complexity introduced by the synthetic trajectories and varied scene conditions. This indicates the dataset's value in facilitating rigorous testing of SLAM systems, which must contend with realistic operational challenges like lighting changes and object rearrangements.

Implications and Future Directions

InteriorNet presents clear practical and theoretical implications for the computer vision community. The dataset's scale and dynamism position it as a critical resource for developing and refining algorithms for SLAM, autonomous navigation, and scene understanding, underpinning more reliable and capable robotic and AR/VR systems.

Looking ahead, further refinement could explore data-driven scene rearrangement and incorporate real-world lighting and scene change ground truths. As the demand for sophisticated indoor environment models grows, InteriorNet offers a robust platform for continued advancements in data-driven spatial perception.

By addressing the limitations of current synthetic datasets and pushing for greater realism, InteriorNet stands as a pivotal resource for the ongoing development of computer vision technologies and applications.