SBA

The RAD (Randomly Assembled Object Construction) dataset is a synthetic 3D LEGO dataset designed for the task of Sequential Brick Assembly (SBA). Here are the key characteristics and details:

High-level explanation:

Consists of 3D objects built with 2x4 LEGO bricks. Objects are randomly assembled following predefined connection configurations. Includes variations like RAD-S and RAD with different numbers of bricks per object. Provides ground truth labels for assembly actions. Includes multi-view images, sequential assembly actions, 3D voxel, conjunction graph, LEGO ldr file, etc.

Motivations and content summary:

Created to provide a large-scale synthetic dataset for training and evaluating SBA models. Allows for controlled experimentation with varying object complexity. RAD-1k: 1000 objects with around 15 bricks each. RAD-S: 10000 objects built with around 20 bricks each. RAD: 10000 objects built with around 60 bricks each. Expands on previous datasets by allowing more flexible brick connections, including bottom-to-top assembly.

Potential use cases:

Pre-training models for sequential brick assembly tasks. Evaluating the scalability of assembly algorithms with increasing object complexity. Studying transfer learning from synthetic to real-world assembly tasks. Benchmarking performance of different SBA approaches. Investigating generalization capabilities of assembly models to different object structures.

The RAD dataset provides a valuable resource for researchers working on 3D object assembly, particularly in developing and testing algorithms that can handle increasingly complex structures built from simple primitives.