ChangeSim

ChangeSim 是一个针对在线场景变化检测 (SCD) 等的数据集。这些数据是在具有环境非目标变化（例如空气混浊度和光照条件变化，以及工业室内环境中的目标对象变化）的照片般逼真的模拟环境中收集的。通过在模拟中收集数据，可以获得多模态传感器数据和精确的地面实况标签，例如 RGB 图像、深度图像、语义分割、变化分割、相机姿势和 3D 重建。虽然之前的在线 SCD 数据集评估模型给定对齐良好的图像对，但 ChangeSim 还提供了原始的未配对序列，这为以端到端方式开发在线 SCD 模型提供了机会，同时考虑了配对和检测。实验表明，即使是最新的基于配对的 SCD 模型也存在配对过程的瓶颈，当环境包含非目标变化时，情况会变得更糟。

ChangeSim is a dataset dedicated to online Scene Change Detection (SCD) and related tasks. The data is collected in photorealistic simulated environments that contain environmental non-target changes, such as variations in air turbidity, lighting conditions, and changes in target objects within industrial indoor environments. By collecting data in simulations, this dataset enables acquisition of multimodal sensor data and accurate ground-truth labels, including RGB images, depth images, semantic segmentation masks, change segmentation masks, camera poses, and 3D reconstructions. While prior online SCD datasets provide well-aligned image pairs for model evaluation, ChangeSim also offers raw unpaired sequences, which enables the development of online SCD models in an end-to-end fashion that accounts for both pairing and detection tasks. Experimental results demonstrate that even state-of-the-art pair-based SCD models encounter bottlenecks in the pairing process, and this issue is further exacerbated when the environment contains non-target changes.

搜集汇总

背景与挑战

背景概述

ChangeSim是一个针对在线场景变化检测的数据集，提供多模态传感器数据和精确的地面实况标签，包括RGB图像、深度图像等，并包含原始的未配对序列，支持端到端的模型开发。数据集主要应用于工业室内环境中的目标对象变化检测。