Robotic Interestingness Dataset (SubTF)
CMU AirLab · 2020 · datasets.bot · datasets.bot page
One-liner. Onboard front-camera video from autonomous ground robots exploring DARPA SubT Challenge mine tunnels, with human interestingness labels for online visual interesting-scene prediction.
Setup
- Datasets / benchmarks: The Robotic Interestingness Dataset (also called the SubT Front Camera, or SubTF, dataset) from CMU's AirLab is a visual scene dataset for studying online prediction of visually 'interesting' scenes for mobile robots. It consists of 7 long-form video sequences (~53-83 minutes each, ~467.7 minutes / ~7.8 hours total) recorded from the front-facing RGB camera of two fully autonomous unmanned ground vehicles (UGVs) operated by CMU Team Explorer, which won first place at the DARPA Subterranean (SubT) Challenge Tunnel Circuit. The robots explored two underground mine tunnels (cumulative 4-8 km linear distance) under challenging GPS- and communication-denied conditions with poor lighting, dripping water, smoke, and cluttered/irregular geometry. Frames are annotated by multiple human labelers for visual interestingness: 15.49% of frames were marked interesting by at least one annotator and 3.58% by at least two. The data is distributed both as processed image sequences (organized as date-ugvN-tunnelM folders, e.g. 0817-ugv0-tunnel0, with ground-truth and train splits) and as raw ROS bag files, hosted via OneDrive/SharePoint links reachable from the AirLab dataset instructions page. The dataset accompanies the paper 'Visual Memorability for Robotic Interestingness via Unsupervised Online Learning' (ECCV 2020 Oral; extended in IEEE Transactions on Robotics 2021), and is supported by the sair-lab/interestingness code (a three-stage long-term/short-term/online learning visual-memory approach) and the wang-chen/SubT data tools (BSD-3-Clause). License: unknown. Download: https://theairlab.org/dataset/interestingness.
- Hardware / simulator: Embodiment: not listed. Environment: industrial. Realness: physical.
Schema
7 front-camera video sequences (~467.7 min total) from two autonomous UGVs in DARPA SubT mine tunnels, provided as processed RGB image folders (date-ugvN-tunnelM) plus ground-truth/train splits and raw ROS bag files; per-frame human interestingness labels.
Links