You will know more about the below 3 things when you finish reading this blog.

1. Current challenges of cost and scalability and how Artificial Intelligence/Machine learning (AI) will help deal with these challenges in 3D content creation?
2. Where do we stand in terms of using AI in our current production pipeline?
3. Which technologies and algorithms we are using?

Today, eCommerce has been incorporating AI into everything such as enhancing the consumer experience, attracting users to accessing their platform, and increasing conversion rates.

Whether you talk about Personalized Recommendations, where Amazon’s “item-to-item collaborative filtering” algorithm is providing tailor-made content direct to the homepage which increased its sales or you talk about voice commands where Voice-activated assistants like Amazon’s Echo, Google Home, and the Apple HomePod, or on mobile devices with Cortana and Siri, AI is ruling the world. AI is a widely used technology in today’s time and industry tycoons like Alibaba, Rakuten, eBay, etc use it for various purposes ranging from product recommendations to fake reviews detection and managing big data. Gartner predicts that 85% of interactions between customers and retailers will happen “without interacting with a human” by 2020.

Despite going out of the way to deliver positive customer experiences, e-commerce stores still struggle to fulfill the product visualization expectations that customers have. It proves to be virtually impossible to offer product representations that are even remotely at par with how they are on-site in physical stores. This is where 3D technology can be leveraged to enable customers to visualize a product from all angles before making a purchase decision. With this, a customer can choose what part of the object they’d like to see, whether they want to zoom in or out the object or they want to rotate the object. All these aims can be realized using 3D modeling capabilities.

3 most common challenges for 3D models every eCommerce player is facing are –

1. Scalability – Everyone wants to have 3D models created at the earliest. We have also heard some clients wanting 1000 models or more in a month. On average one model creation time including material is 10-12 hours so we are talking about 12,000 man-hours of work which need 80 people in a month. Imagine how difficult it is to scale up so quickly and set up a team that can deliver 1000 or more models.
2. Cost – 3D is an additional cost as it’s seen right now by decision-makers. Executives have a lot of pressure to reduce the cost further down.
3. Quality – Quality is important more than anything else, you won’t have realistic 3D models with the right dimension and material.

SuperDNA realized this challenge earlier in the cycle and directly focused on AI experiments to integrate machine learning in creating 3D assets. We are investing in research and also creating our own AI/Machine Learning utilities which can help us to address these challenges. Below are some of the initiatives that we took –

1. Repurposing – While working for large clients we are having separate production pipelines. We understand that after developing repetitive models and materials, we can use existing models and textures to bring the cost down. We have our asset libraries as well which we have purchased which have more than 50,000 models and textures. We run our machine learning recommendation engine to provide the best match to our 3D modeling and 3D texture artist so that they don’t need to start from scratch. Also, preapproved models are already of high quality, so we face fewer quality issues.

Current Status of SuperDNA3D Lab – We started this in early 2018 and this utility is quite mature. We can save as much as 15% on efforts/costs which we pass on to our customers.

Technology background – We are using veggie, a CNN trained using Caffe deep learning framework. When you pass any image through the network it gives you pattern and shape features at the end of convolutional and fully connected (FC) layers, respectively. Those features are compressed using a compression matrix and then we use the nearest neighbor classifier to find the best match. The pattern and shape features of the database are pre-computed and saved in CSV files.

1. Creation of textures/materials and 3D models.

3D texture creation –
We also invested in scratch by creating low-resolution textures directly to the high-resolution texture which means if we have scanned resolution textures then we create directly high-resolution textures from our utilities. There are a lot of tools available online but they miss out on details and that’s what we are trying to address here.

3D model creation –
We are in the early stage of creating models from scratch for 3D directly from 2D images. We think if do a breakthrough in this utility then we will be able to solve most of the problems. Human intervention will still be needed but it will be a huge cost saving.

Technology Solution background –

SuperDNA is using an interesting approach to solve this problem. We came up with the idea of combining the research done by people with the traditional approach to solving any computer vision problem. Utilizing the research from Wu and Hane’s papers, we created 2 neural networks, first, one will be used to do the feature extraction and another network will predict the coordinate based on the features provided to this network. This approach is unique since we are combining 2 power networks, i.e., a Convolutional Neural Network(CNN) for feature extraction and a recurrent neural network for predicting coordinates of the 3D object.

How we are doing it?

1. We are trying to eat the elephant bit by bit.
2. We have very well integrated the AI utilities within our workflows so we test our products and keep making them better every day.

In a nutshell –

To conclude, Integrating AI usage into 3D asset creation is a daunting task but the results are positive. By working on re-purposing and creation separately we can address challenges of scalability, quality, and costs.

Feel free to reach out to me if you want to know more …, Click here