CoolFace
Datasetpublic

Nanase1234/scannet-intention-10k

ScanNet Intention Dataset Indoor scene images with human intention annotations for object interaction. Dataset Description This dataset is derived from ScanNet indoor scene scans, with GPT-generated intention annotations. Each sample contains an indoor scene image, a target object, and 3 plausible human intentions. Splits Split Scenes Images/Queries train 596 8,853 test 100 1,855 Note: Train and test splits are separated by… See the full description on the dataset page: https://huggingface.co/datasets/Nanase1234/scannet-intention-10k.

sourceHugging Facemitupdated 9mo agoView on Hugging Face
0likes37downloads
Dataset Card

ScanNet Intention Dataset

Indoor scene images with human intention annotations for object interaction.

Dataset Description

This dataset is derived from ScanNet indoor scene scans, with GPT-generated intention annotations. Each sample contains an indoor scene image, a target object, and 3 plausible human intentions.

Splits

SplitScenesImages/Queries
train5968,853
test1001,855

Note: Train and test splits are separated by physical scene to prevent data leakage.

Fields

FieldDescription
idSample ID
imageIndoor scene image
scenePhysical scene ID (e.g., "scene0000")
scanScan session ID (e.g., "scene0000_00")
target_categoryTarget object category (e.g., "chair", "table")
bboxBounding box [x, y, w, h] in COCO format
scene_reasoningScene context description
intention_1/2/3Three plausible interaction intentions

Quick Start

python
from datasets import load_dataset

ds = load_dataset("Nanase1234/scannet-intention-10k")

# Access train split
train_sample = ds["train"][0]
print(train_sample["target_category"])
print(train_sample["intention_1"])

# Access test split
test_sample = ds["test"][0]
print(test_sample["scene"])

Statistics

  • —Total Samples: 10,708
  • —Physical Scenes: 696
  • —Object Categories: 34
  • —Train/Test Split: Scene-level (no scene overlap)

BBox Format

COCO format: [x, y, width, height]

  • —x, y: Top-left corner coordinates (pixels)
  • —width, height: Box dimensions (pixels)