[Paper Review] Product Aesthetic Design: A Machine Learning Augmentation

Alex Burnap, John R. Hauser+1 more|arXiv (Cornell University)|Jul 17, 2019

Aesthetic Perception and Analysis22 citations

TL;DR

This paper proposes a hybrid machine learning model combining variational autoencoders (VAE), generative adversarial networks (GAN), and supervised learning to predict aesthetic appeal and generate innovative product designs. Trained on 203 SUVs and 180,000 images, the model improves aesthetic prediction by 43.5% over baseline and generates designs that are both consumer-appealing and consistent with real-world market trends from five years later.

ABSTRACT

Aesthetics are critically important to market acceptance. In the automotive industry, an improved aesthetic design can boost sales by 30% or more. Firms invest heavily in designing and testing aesthetics. A single automotive "theme clinic" can cost over $100,000, and hundreds are conducted annually. We propose a model to augment the commonly-used aesthetic design process by predicting aesthetic scores and automatically generating innovative and appealing product designs. The model combines a probabilistic variational autoencoder (VAE) with adversarial components from generative adversarial networks (GAN) and a supervised learning component. We train and evaluate the model with data from an automotive partner-images of 203 SUVs evaluated by targeted consumers and 180,000 high-quality unrated images. Our model predicts well the appeal of new aesthetic designs-43.5% improvement relative to a uniform baseline and substantial improvement over conventional machine learning models and pretrained deep neural networks. New automotive designs are generated in a controllable manner for use by design teams. We empirically verify that automatically generated designs are (1) appealing to consumers and (2) resemble designs which were introduced to the market five years after our data were collected. We provide an additional proof-of-concept application using opensource images of dining room chairs.

Motivation & Objective

• To address the high cost and inefficiency of traditional aesthetic design processes in industries like automotive, where theme clinics exceed $100,000 per session.
• To develop a data-driven, scalable method for predicting consumer aesthetic preferences in product design.
• To generate novel, controllable, and appealing product designs automatically, reducing reliance on costly human-led design iterations.
• To validate that generated designs are not only appealing but also align with real-world market introductions over time.
• To extend the model’s applicability beyond automobiles using a proof-of-concept on dining room chairs.

Proposed method

• The model integrates a probabilistic variational autoencoder (VAE) for learning a disentangled latent representation of product images.
• Adversarial training components from generative adversarial networks (GAN) are used to improve the realism and diversity of generated designs.
• A supervised learning head is trained on consumer-rated aesthetic scores to predict perceived appeal from image features.
• The model enables conditional generation of new designs by conditioning on latent codes that control specific aesthetic attributes.
• Training is performed on a dataset of 203 SUVs with consumer ratings and 180,000 high-quality, unrated product images.
• The framework supports controllable generation, allowing design teams to explore variations based on desired aesthetic directions.

Experimental results

Research questions

• RQ1 Can a hybrid deep learning model effectively predict consumer aesthetic preferences for product designs with high accuracy?
• RQ2 To what extent can the model generate novel, visually appealing, and diverse product designs that resemble real-world market introductions?
• RQ3 How does the model’s predictive performance compare to conventional machine learning models and pretrained deep networks in aesthetic scoring?
• RQ4 Can the generated designs maintain aesthetic quality while being controllable and aligned with human design intent?
• RQ5 Does the model generalize to other product categories beyond automobiles, as demonstrated by the dining chair proof-of-concept?

Key findings

• The model achieves a 43.5% improvement in aesthetic prediction accuracy over a uniform baseline, significantly outperforming conventional machine learning models.
• The model demonstrates substantial improvement over both standard machine learning models and pretrained deep neural networks in predicting consumer appeal.
• Generated designs are empirically validated as appealing to consumers, with strong alignment to real-world market introductions from five years after the training data was collected.
• The model successfully generates diverse, high-quality designs in a controllable manner, enabling practical use by design teams.
• The proof-of-concept on dining room chairs confirms the model’s generalizability to other product categories beyond automobiles.
• The integration of VAE, GAN, and supervised learning components results in a robust architecture capable of both accurate prediction and meaningful generation.