BIKED++

A Multimodal Dataset of 1.4 Million Bicycle Image and Parametric CAD Designs.

Lyle Regenwetter1, Yazan Abu Obaideh2, Amin Heyrani Nobari1, Faez Ahmed1

1MIT  2Sigasi 

Explore

We introduce BIKED++, an addition to the BIKED family supporting large-scale multimodal capabilities.

Code
Paper
BIKED
Download

Overview

BIKED ++ introduces an image rendering pipeline based on BikeCAD. The render allows users to rapidly generate images from BIKED-style parametric bike vectors.

Using the renderer, we generate a dataset of 1.4 million bicycle parametric-image pairs. We additionally calculate CLIP embeddings for each image in the dataset.

Fast and Accurate Embedding Calcualtion

BIKED ++ is large enough to train effective direct embedding models. We train a residual network to estimate CLIP embeddings directly from parametric vectors.

This model is fast and accurate, allowing for near-instant similarity calculations to reference text or images.

This predictive model enables many of the multimodal features in the MCD project.

Citations

Chicago

Regenwetter, Lyle, Yazan Abu Obaideh, Amin Heyrani Nobari, and Faez Ahmed. 'BIKED++: A Multimodal Dataset of 1.4 Million Bicycle Image and Parametric CAD Designs.' arXiv preprint arXiv:2402.05301 (2024).

Bibtex

@article{regenwetter2024biked++,
     title= {BIKED++: A Multimodal Dataset of 1.4 Million Bicycle Image and Parametric CAD Designs},
     author={Regenwetter,
     Lyle and Obaideh,
     Yazan Abu and Nobari,
     Amin Heyrani and Ahmed,
     Faez},
     journal={arXiv preprint arXiv:2402.05301},
     year={2024}}