Wire Detection Dataset

video · The Wire Detection Dataset accompanies the IROS 2017 paper 'Wire Detection using Synthetic Data and Dilated Convolutional Networks for Unmanned Aerial Vehicles' (Ratnesh Madaan, Daniel Maturana, Sebastian Scherer, Robotics Institute, Carnegie Mellon University / AirLab). The method targets detection of thin structures such as power lines from unmanned aerial vehicles, and was trained purely on synthetic imagery and evaluated on real video — a sim-to-real transfer result. **What actually ships.** The download (wire-detection-dataset.tar.gz, 1.7 GB via Google Drive) contains a single top-level directory, wires_usf/, holding 3,499 sample folders. Every folder is named as a frame of a source video from the University of South Florida wire/thin-object video collection — the 'candamo.*', 'DSCN*', 'Temp*' and 38 'uvs*' series — so the samples come in long runs of consecutive frames from the same clip rather than as independent stills. The synthetic training images the paper describes are not part of this archive; what is released is the labelled real-video evaluation set. Cameras vary with the source clip: some series are colour ground-level video of power lines against sky, others are low-resolution grayscale. **Per-sample layout.** Each folder contains original_image.png (the frame), labeled_ground_truth.png (pixel labels, 1 = non-wire, 2 = wire), ground_truth_viz.png (a black-and-white visualisation of those labels), and labels.ground (a text file giving the endpoint coordinates of each labelled line segment). Distributed under CC-BY-4.0.

The Wire Detection Dataset accompanies the IROS 2017 paper 'Wire Detection using Synthetic Data and Dilated Convolutional Networks for Unmanned Aerial Vehicles' (Ratnesh Madaan, Daniel Maturana, Sebastian Scherer, Robotics Institute, Carnegie Mellon University / AirLab). The method targets detection of thin structures such as power lines from unmanned aerial vehicles, and was trained purely on synthetic imagery and evaluated on real video — a sim-to-real transfer result. **What actually ships.** The download (wire-detection-dataset.tar.gz, 1.7 GB via Google Drive) contains a single top-level directory, wires_usf/, holding 3,499 sample folders. Every folder is named as a frame of a source video from the University of South Florida wire/thin-object video collection — the 'candamo.*', 'DSCN*', 'Temp*' and 38 'uvs*' series — so the samples come in long runs of consecutive frames from the same clip rather than as independent stills. The synthetic training images the paper describes are not part of this archive; what is released is the labelled real-video evaluation set. Cameras vary with the source clip: some series are colour ground-level video of power lines against sky, others are low-resolution grayscale. **Per-sample layout.** Each folder contains original_image.png (the frame), labeled_ground_truth.png (pixel labels, 1 = non-wire, 2 = wire), ground_truth_viz.png (a black-and-white visualisation of those labels), and labels.ground (a text file giving the endpoint coordinates of each labelled line segment). Distributed under CC-BY-4.0.

The Wire Detection Dataset accompanies the IROS 2017 paper 'Wire Detection using Synthetic Data and Dilated Convolutional Networks for Unmanned Aerial Vehicles' (Ratnesh Madaan, Daniel Maturana, Sebastian Scherer, Robotics Institute, Carnegie Mellon University / AirLab). The method targets detection of thin structures such as power lines from unmanned aerial vehicles, and was trained purely on synthetic imagery and evaluated on real video — a sim-to-real transfer result. **What actually ships.** The download (wire-detection-dataset.tar.gz, 1.7 GB via Google Drive) contains a single top-level directory, wires_usf/, holding 3,499 sample folders. Every folder is named as a frame of a source video from the University of South Florida wire/thin-object video collection — the 'candamo.*', 'DSCN*', 'Temp*' and 38 'uvs*' series — so the samples come in long runs of consecutive frames from the same clip rather than as independent stills. The synthetic training images the paper describes are not part of this archive; what is released is the labelled real-video evaluation set. Cameras vary with the source clip: some series are colour ground-level video of power lines against sky, others are low-resolution grayscale. **Per-sample layout.** Each folder contains original_image.png (the frame), labeled_ground_truth.png (pixel labels, 1 = non-wire, 2 = wire), ground_truth_viz.png (a black-and-white visualisation of those labels), and labels.ground (a text file giving the endpoint coordinates of each labelled line segment). Distributed under CC-BY-4.0.

simulation · The Wire Detection Dataset accompanies the IROS 2017 paper 'Wire Detection using Synthetic Data and Dilated Convolutional Networks for Unmanned Aerial Vehicles' (Ratnesh Madaan, Daniel Maturana, Sebastian Scherer, Robotics Institute, Carnegie Mellon University / AirLab). The method targets detection of thin structures such as power lines from unmanned aerial vehicles, and was trained purely on synthetic imagery and evaluated on real video — a sim-to-real transfer result. **What actually ships.** The download (wire-detection-dataset.tar.gz, 1.7 GB via Google Drive) contains a single top-level directory, wires_usf/, holding 3,499 sample folders. Every folder is named as a frame of a source video from the University of South Florida wire/thin-object video collection — the 'candamo.*', 'DSCN*', 'Temp*' and 38 'uvs*' series — so the samples come in long runs of consecutive frames from the same clip rather than as independent stills. The synthetic training images the paper describes are not part of this archive; what is released is the labelled real-video evaluation set. Cameras vary with the source clip: some series are colour ground-level video of power lines against sky, others are low-resolution grayscale. **Per-sample layout.** Each folder contains original_image.png (the frame), labeled_ground_truth.png (pixel labels, 1 = non-wire, 2 = wire), ground_truth_viz.png (a black-and-white visualisation of those labels), and labels.ground (a text file giving the endpoint coordinates of each labelled line segment). Distributed under CC-BY-4.0.

Open · CC-BY-4.0 · custom