--- license: mit dataset_info: features: - name: scene_id dtype: string - name: ann_id dtype: string - name: object_name dtype: string - name: object_ids list: int64 - name: description dtype: string - name: num_attributes_target dtype: int64 - name: num_attributes_anchors dtype: int64 - name: num_relationships_target dtype: int64 - name: num_relationships_anchors dtype: int64 - name: granularity dtype: string - name: attribute_type_color dtype: bool - name: attribute_type_size dtype: bool - name: attribute_type_shape dtype: bool - name: attribute_type_number dtype: bool - name: attribute_type_material dtype: bool - name: attribute_type_texture dtype: bool - name: attribute_type_function dtype: bool - name: attribute_type_style dtype: bool - name: attribute_type_text_label dtype: bool - name: attribute_type_state dtype: bool - name: relationship_type_near dtype: bool - name: relationship_type_far dtype: bool - name: relationship_type_viewpoint_dependent dtype: bool - name: relationship_type_vertical dtype: bool - name: relationship_type_contain dtype: bool - name: relationship_type_arrangement dtype: bool - name: relationship_type_ordinal dtype: bool - name: relationship_type_comparison dtype: bool - name: coreferences dtype: bool - name: target_not_first_np dtype: bool - name: anchor_type_single dtype: bool - name: anchor_type_multiple dtype: bool - name: anchor_type_non_object dtype: bool - name: anchor_type_viewpoint dtype: bool - name: negation dtype: bool splits: - name: validation num_bytes: 72126 num_examples: 349 download_size: 45798 dataset_size: 72126 configs: - config_name: default data_files: - split: validation path: data/validation-* --- # ViGiL3D Dataset This repository contains the ViGiL3D benchmark dataset, introduced in [ViGiL3D: A Linguistically Diverse Dataset for 3D Visual Grounding](https://3dlg-hcvc.github.io/vigil3d/). Prompts for both ScanNet and ScanNet++ are included in the same split. ## Dataset Attributes The key attributes are as follows: **scene_id: (str)** the scene identifier as defined in [ScanNetv2](http://www.scan-net.org/) and [ScanNet++](https://kaldir.vc.in.tum.de/scannetpp/). \ **ann_id: (int)** the global annotation ID. \ **object_name: (str)** the name of the target object(s). \ **description: (str)** the language description of the target object(s) in the scene. \ **object_ids: (List[int])** the target object ID(s) as defined in [ScanNetv2](http://www.scan-net.org/). In addition, we include the following metadata attributes characterizing the language patterns present in each description: **num_attributes_target (int)** number of attributes describing the target object in the description \ **num_attributes_anchors (int)** number of attributes describing anchor objects in the description \ **num_relationships_target (int)** number of relationships describing the target object in the description \ **num_relationships_anchors (int)** number of relationships describing only anchor objects in the description \ **granularity (string)** the specificity of the noun used to refer to the target (`generic`, `categorical`, or `fine-grained`) \ **attribute_type_color (bool)** true if the color of an object is included in the description \ **attribute_type_size (bool)** true if the size of an object is included in the description \ **attribute_type_shape (bool)** true if the shape of an object is included in the description \ **attribute_type_number (bool)** true if the number of an object is included in the description \ **attribute_type_material (bool)** true if the material of an object is included in the description \ **attribute_type_texture (bool)** true if the texture of an object is included in the description \ **attribute_type_function (bool)** true if the function of an object is included in the description \ **attribute_type_style (bool)** true if the style of an object is included in the description \ **attribute_type_text_label (bool)** true if a text label on an object is included in the description \ **attribute_type_state (bool)** true if the state of an object (e.g. `open`, `closed`) is included in the description \ **relationship_type_near (bool)** true if an object is described as _near_ another in the description \ **relationship_type_far (bool)** true if an object is described as _far_ from another in the description \ **relationship_type_viewpoint_dependent (bool)** true if an object is described directionally relative to another (e.g. `left`, `right`) in the description \ **relationship_type_vertical (bool)** true if an object is described above, below, or on top of another in the description \ **relationship_type_contain (bool)** true if an object is described within or inside of another in the description \ **relationship_type_arrangement (bool)** true if an object is described as part of a larger arrangement (e.g. `row`, `stack`) in the description \ **relationship_type_ordinal (bool)** true if an object is described with an ordinal position (e.g. `first`, `third`) in the description \ **relationship_type_comparison (bool)** true if an object is compared to others with respect to an attribute (e.g. taller, widest) in the description \ **coreferences (bool)** true if an object is named using a coreference (e.g. `it`) in the description \ **target_not_first_np (bool)** true if the target object is not the first noun phrase in the description \ **anchor_type_single (bool)** true if one of the anchor objects in the description is a single object \ **anchor_type_multiple (bool)** true if one of the anchor objects in the description refers to multiple objects \ **anchor_type_non_object (bool)** true if one of the anchor objects in the description refers to a non-object (e.g. a region of the room) \ **anchor_type_viewpoint (bool)** true if one of the anchor objects in the description refers to an agent \ **negation (bool)** true if the description incldues negation \ ## Citation ```bibtex @inproceedings{wang2025vigil3d, author={Wang, Austin T and Gong, ZeMing and Chang, Angel X}, title={ViGiL3D: A linguistically diverse dataset for 3d visual grounding}, booktitle={Proceedings of the Conference of the Association for Computational Linguistics (ACL)}, month={July}, year={2025}, } ```