IRIS: Industrial Room In Saclay

IRIS is a multimodal dataset that provides dense station-based LiDAR point clouds, panoramic images, a CAD model, a piping and instrumentation diagram (P&ID), annotated bounding boxes, and segmentation masks for several object categories. The scene represents a large and complex industrial room covering 530 m². It contains objects of various shapes, colors, and sizes such as pipes, valves, pumps and gauges.

This dataset has been used for 3D point visibility estimation and scene-functional alignment, but it is also suitable for many other computer vision applications.

More details are available on the dataset page.

IRIS data

All data modalities are aligned within the same 3D coordinate system.

IRIS-v1

point_cloud_stations_01 to 14.zip: The raw point cloud of the entire scene with very high density from Zoller+Frölich 5016C device. In total 67 files, one per station acquisition, and more than 2.1 billion points.

point_cloud_stations_info.zip: The stations information. It provides the position in the scene (translation vector), and orientation (rotation vector and angle).

point_cloud_stations_merged_zoneA to C_density_10%.ply: Subparts covering in total 30% of the whole scene. Multiple station clouds have been merged. The density is reduced by a factor of 10 compared to the original scans. The resulting clouds are still dense.

IRIS-v2

panoramic images:

XPhase Pro X2 camera:

images_xphase_1 to 4.zip: 312 images with a resolution of 16384x8192 pixels.

images_xphase_camera_info.zip: The camera positions in the 3D scene (translation vector), and orientation (rotation vector and angle).

Zoller+Frölich 5016C camera:

images_ZF.zip: 67 images with a resolution of 10000x5000 pixels.

images_ZF_camera_info.zip: The camera positions in the 3D scene (translation vector), and orientation (rotation vector and angle).

CAD model:

cad_model.fbx: A CAD model reconstructed semi-automatically to closely match the point cloud.

cad_model_object_tree.txt: Hierarchical organization of the CAD model. Level 1: object categories in English. Level 2: object names in French. 

cad_model_extracted_mesh.ply: Mesh extracted from the CAD model.

segmentation masks from the CAD model: 

annotations_segmentation_cubemaps_xphase.zip: Cubemaps extracted from the XPhase Pro X2 images together with 24549 COCO-format segmentation masks distributed across 16 categories. Annotations are projected from the CAD model. Categories correspond to those defined in cad_model_object_tree.txt.

annotated boxes:

annotations_boxes_cubemaps_ZF.zip: Cubemaps extracted from the Zoller+Frölich 5016C images together with 6303 COCO-format boxes distributed across 171 categories. The boxes are human annotated.

P&ID (pipe and instrumentation diagram):

piping_and_instrumentation_diagram.pdf: Complete piping and instrumentation diagram in PDF-image format.

IRIS-VIS data

IRIS-VIS is a dataset specifically designed for the point visibility estimation task. iris-vis.zip contains all the data. More details are available on the IRIS-VIS paper page.

The folder show contains:

An example of the visibility ground truth (see paper for the ground truth construction).

An example of the complex visibility areas (see paper).

Visualizations of the corrected mesh from the CAD model on a pipe and a valve (see paper).

ZoneA, B and C are the same scenes as in IRIS. They are used for quantitative experiments. Each scene includes:

3 viewpoints,

for two densities (2% and 10% of the raw point cloud):

The input cloud (paired with the cad model).

For each viewpoint: The visibility ground truth cloud and indices (computed from the cad model).

For each viewpoint: The cloud and indices of the complex areas (computed from the ground truth).

ZoneA.1 and A.2 are two subscenes of the zoneA used for the qualitative visualizations. They include:

1 viewpoint (the same for A.1 and A.2).

The same data as for the scenes A, B and C (see above), including the raw and input clouds for the visibility estimation task.

The visibility prediction cloud and indices for each method.

The visibility evaluation cloud for each method: The prediction cloud merged with the ground truth cloud and colorized according to true positive (blue), false positive (purple), false negative (orange).

The camera parameters used for Vis2mesh.