SG-FRONT
SG-FRONT
SG-FRONT provides semantic scene graphs for 3D-FRONT (refined version). A semantic scene graph is defined by a set of tuples between nodes and edges where nodes represent specific 3D object instances. Nodes are defined by their semantics, a hierarchy of classes. The edges in our graphs are the semantic relationships (predicates) between the nodes, including standing on or same material as, following 3DSSG.
Data Organization
Download link: SG-FRONT and additional bboxes. Download and unzip them to path/to/3D-FRONT. The data in the provided files is organized as follows:
SG-FRONT.zip
|-- relationships_{type}_trainval.json
List of all triplets consisting of semantic classes (nodes) and semantic relationships (edges) in the trainval split.
|-- relationships_{type}_test.json
List of all triplets consisting of semantic classes (nodes) and semantic relationships (edges) in the test split.
|-- classes_{type}.txt
List of the semantic classes in the current room type.
|-- relationships.txt
List of all relationship types.
bbox.zip
|-- obj_boxes_{type}_trainval.json
Oriented bounding boxes in each scene in the trainval split.
|-- obj_boxes_{type}_test.json
Oriented bounding boxes in each scene in the test split.
|-- boxes_centered_stats_{}_trainval.txt
Box normalization parameters in the trainval split.
|-- boxes_centered_stats_{}_test.txt
Box normalization parameters in the test split.
|-- cat_jid_*.json (not necessary for commonscenes)
Object instance id with the bounding box size.
Data Formats
SG-FRONT is for the general purpose of scene graph research:
Object relationships (edges) (relationships_*.json):
{
"scans": [
{
"scan": "MasterBedroom-33296",
"objects": {"1": "dressing_table", "2": "nightstand", "3": "stool", "4": "double_bed", "5": "cabinet", "6": "pendant_lamp", "7": "wardrobe", "8": "floor"},
"relationships": [
[
6, // subject instance id in the current scene
5, // object instance id in the current scene
1, // relationships id (order aligned with relationships.txt)
"left" // textual semantic relationship (from relationships.txt)
],
[ 5, 2, 14, "same super category as" ],
[ 4, 1, 10, "taller than" ],
[ 7, 2, 15, "same material as" ],
...
],
}, {
...
}
]
}
Bbox information is for additional usages:
Object bboxes (obj_boxes_*.json):
{
"MasterBedroom-33296": { // room id from preprocessed 3D-FRONT
"1": {"param7": [1.0406779999999998, 1.098808, 0.4491219999999996, -5.370314764104733, 0.0, -4.066067038275899, 0], // bbox size and pose (l,h,w,x,y,z,angle), with positive y-axis up. Notice that bbox size is different from ATISS's definition.
"8points": [[-5.890653764104733, 0.0, -4.290628038275899], [-5.890653764104733, 0.0, -3.841506038275899], [-5.890653764104733, 1.098808, -4.290628038275899], [-5.890653764104733, 1.098808, -3.841506038275899], [-4.849975764104733, 0.0, -4.290628038275899], [-4.849975764104733, 0.0, -3.841506038275899], [-4.849975764104733, 1.098808, -4.290628038275899], [-4.849975764104733, 1.098808, -3.841506038275899]], // bbox vertices
"scale": [1, 1, 1], // always 1
"model_path": "/media/ymxlzgy/Data/Dataset/3D-FRONT/3D-FUTURE-model/1791e0c9-15fc-4fa9-80df-e73332ed6ce0/raw_model.obj"}, // the object model in 3D-FUTURE. The head path needs to be changed.
"2": {"param7": [0.48633499999999996, 0.607177, 0.4845579999999998, -2.560559506120386, 0.0, -4.061041670654544, 0],
"8points": [[-2.803745506120386, 0.0, -4.303502670654543], [-2.803745506120386, 0.0, -3.8189446706545436], [-2.803745506120386, 0.607177, -4.303502670654543], [-2.803745506120386, 0.607177, -3.8189446706545436], [-2.3174105061203862, 0.0, -4.303502670654543], [-2.3174105061203862, 0.0, -3.8189446706545436], [-2.3174105061203862, 0.607177, -4.303502670654543], [-2.3174105061203862, 0.607177, -3.8189446706545436]],
"scale": [1, 1, 1],
"model_path": "/media/ymxlzgy/Data/Dataset/3D-FRONT/3D-FUTURE-model/62319141-32fa-4353-b2ff-ec31ca232e3e/raw_model.obj"},
...,
"scene_center": [-2.9265, 0.0, -2.7]}, // the average center of the current room
"MasterBedroom-9404": {
...,
"scene_center": [2.96192, 0.0, 0.32996000000000003]}
...
}
Notes:
The dataset is released under MIT license. Our automatic labeling setup can be downloaded here for reference and potential adjustment. This can ONLY be run after preprocessing 3D-FRONT following Diffuscene or ATISS.