InfiniBench: Infinite Benchmarking for Visual Spatial Reasoning with Customizable Scene Complexity

Dataset Summary

InfiniBench is a dataset generator for procedurally generated 3D indoor environments designed for multimodal AI, embodied agents, and spatial reasoning tasks.

These scenes were created using an enhanced version of Infinigen that utilizes Agentic Constraint Generation (LLM-driven scene descriptions and code refinement) and Cluster-Based Layout Optimization to create logical, rigid-body furniture arrangements.

Each entry in this dataset represents a unique scene containing the raw 3D assets (.blend), optimized camera navigation trajectories (frontier-based exploration), rendered video feeds, and rich metadata for Visual Question Answering (VQA).

Supported Tasks

This dataset supports a variety of tasks described in the associated framework:

• Visual Question Answering (VQA): Measurement, perspective-taking, and spatiotemporal ordering tasks.

Dataset Structure

The dataset is organized by Scene ID. Each folder contains the assets, intermediate navigation results, renders, and metadata for a single generated environment.

Folder Layout

dataset_root/
├── scene_ID/
│   ├── scene.blend                     # Full 3D environment (Blender)
│   ├── trajectory/
│   │   ├── trajectory_data.csv           # Position & Rotation (Euler) for every frame
│   │   └── segmentation_maps/          # 2D Segmentation masks corresponding to the path
│   ├── render/
│   │   ├── frames/                     # Individual rendered frames
│   │   └── trajectory_video.mp4        # Compiled video of the optimized path
│   └── metadata/
│       ├── object_bbox_dimensions.csv  # 3D Bounding box sizes for all objects
│       ├── object_appearance.csv       # Order of appearance in the video