Creating an Apparel Detection App with NVIDIA GPU-Accelerated AI Technologies

People with pieces of clothes labeled by AI.

Vision AI solution provider Drishtic AI developed an apparel detection application using NVIDIA TAO Toolkit and DeepStream SDK.

This guest post was submitted by Drishtic AI Lead Developer, Priti Gavali and Technical Architect, Archana Borawake.

The fashion industry is seeing many changes in terms of new technologies and evolving consumer trends. As one of the fastest-growing sectors in retail, the fashion industry is using data to better understand consumer’s clothing tastes and preferences. Drishtic AI’s solution uses computer vision to analyze photos and videos, deciphers the most trendy styles, and compile data into a useful format for customers. 

Apparel companies can use this data to assess demand and develop clothing that appeals to individuals based on their age, gender, and preferences. By producing styles that are current and popular, clothing companies can reduce waste and create a more sustainable industry.

Drishtic AI uses the NVIDIA Metropolis platform to develop AI-enabled video analytics applications. Drishtic AI’s goal is to use advanced tools and a full-stack approach to achieve faster development times and create extremely compute-optimized solutions for developing vision AI applications. To create its apparel detection application, Drishtic AI used a technology stack built on the NVIDIA EGX platform, incorporating NVIDIA-Certified Systems with NVIDIA GPUs.

The Drishtic AI team started by building the dataset primarily from open source images and videos, using 10400 images with 800 images per clothing class. It annotated all of the images for 13 different classes of clothing—t-shirt, shirt, shorts, skirt, top, dress, pants, denim, hoodie, jacket, cardigan, jumpsuit, and sweater. They used Detectnet_V2 pre-trained models combined with TAO toolkit to shorten their AI development pipeline. To deploy the application, the team used NVIDIA Deepstream SDK for processing raw videos, optimizing video decoding, and accelerating image transformations in real time.

The following videos demonstrate how the application can identify the types of garments people are wearing, even as they move or shift between different light settings. This deep learning model is critical for helping retailers and apparel manufacturers get real-world feedback on current and emerging fashion trends.

Check out the step-by-step approach to develop the apparel detection application in the Drishtic/ApparelDetection GitHub repo.

Source:: NVIDIA