THEODORE (Learning from THEODORE)

最近关于从透视图中合成室内数据集的工作表明，使用卷积神经网络 (CNN) 可以显着改善目标检测结果。在本文中，我们介绍了 THEODORE：一个新颖的大型室内数据集，包含 14 个类别的 100,000 张高分辨率多样化鱼眼图像。为此，我们创建了客厅、不同人物角色和室内纹理的 3D 虚拟环境。除了从虚拟环境中捕获鱼眼图像外，我们还为语义分割、实例掩码和对象检测任务的边界框创建注释。我们将我们的合成数据集与全向图像的最先进的真实世界数据集进行比较。基于 MS COCO 权重，我们表明我们的数据集非常适合微调 CNN 以进行目标检测。通过图像合成和域随机化对我们的模型进行高度概括，我们在高清分析数据集上的班级人员的 AP 高达 0.84。

Recent research on synthetic indoor datasets generated from perspective views has shown that convolutional neural networks (CNNs) can markedly improve object detection performance. In this paper, we introduce THEODORE: a novel large-scale indoor dataset containing 100,000 high-resolution diverse fisheye images across 14 categories. To create this dataset, we developed 3D virtual environments of living rooms with various human characters and indoor textures. Beyond capturing fisheye images from these virtual environments, we also generated annotations including bounding boxes for semantic segmentation, instance masking, and object detection tasks. We compare our synthetic dataset against state-of-the-art real-world omnidirectional image datasets. Using pre-trained weights from MS COCO, we demonstrate that our dataset is highly suitable for fine-tuning CNNs for object detection. Through image synthesis and domain randomization to achieve strong model generalization, we achieve an average precision (AP) of up to 0.84 for the person class on the HD Analytics Dataset.

搜集汇总

背景与挑战

背景概述

THEODORE是一个包含10万张高分辨率鱼眼图像的合成数据集，涵盖14个室内场景类别，提供语义分割、实例掩码和对象检测标注。该数据集通过3D虚拟环境生成，适合用于CNN模型微调，在目标检测任务中表现出色（AP高达0.84）。