TechCompass #86: Generative AI - Computer Vision

Recent years saw a remarkable shift in computer vision, from traditional, CNN-only to transformer-based models, with foundational, diffusion, and language-guided methodologies.

Advancements in image synthesis, video processing, 3D reconstruction, and multimodality signal a transformative phase with implications for robotics. Anticipated breakthroughs extend to edge computing, driven by drones, UAVs, IoT, and the demand for lighter models in remote locations.

Major chip manufacturers and hyperscalers lead in developing specialized AI cores, mirrored in the proliferation of dedicated model optimization packages. This transformative wave not only shapes research but also increases practical applications across diverse domains, underlining the field’s adaptability and potential to redefine machine perception.

Trend 1: Compact multimodal intelligence revolutionizes industries

Retail, logistics, healthcare, and manufacturing currently harness pretrained foundation models for computer vision tasks such as image classification, object detection, and segmentation. While these models offer rapid customization, they are large and often require substantial data for fine tuning.

Compact, task-agnostic models are poised to replace their data-hungry counterparts. They promise faster adaptation, increased accuracy, and less reliance on extensive data, making AI solutions more accessible and efficient across industries.

Integrating computer vision with other modalities, such as language processing, opens new horizons. Merging computer vision with robotics and human interactions holds potential to revolutionize healthcare, autonomous vehicles, and manufacturing, creating intelligent systems that redefine industry standards and enhance our daily lives.

A US telecom giant partnered with Infosys to create an advanced object detection model on Android devices using computer vision. It enabled field engineers to efficiently evaluate installation or repair tasks, saving $150,000 annually on repairs and gaining 900 hours per year. The optimized operational expenses and improved customer experience.

Trend 2: Key point detection elevates retail experience

Key point detection identifies and localizes specific points of interest in an image, including body pose, to analyze ongoing human behavior. This form of computer vision can be used by businesses to analyze interactions with products and provide actionable insights for improved customer engagement.?

Current trends embrace convolution-based models, utilizing both top-down and bottom-up approaches. Despite challenges with viewing different angles in training datasets, the forthcoming integration of next-generation foundation models promises to increase accuracy further. Advancements in the field are already being used to enhance ergonomic assessments in health and safety, transforming interactive gaming experiences. Future advancements will include integrating innovative algorithms such as ‘track anything’ and time-series forecasting, promising further accuracy in key point detection. Combining generative AI with these forward-looking algorithms promises a future with even more refined precision for human activity analysis, opening new avenues in health, safety, gaming, and more. As industries continue to harness computer vision technologies, this synergy of algorithms promises to redefine standards and applications, paving the way for a more accurate and versatile key point detection landscape.

Infosys’ Retail Lab employs advanced body pose key point detection for a seamless shopping experience. By identifying key body movements like elbows, wrists, and fingers, firms gain in-depth insights into customer behavior and product interactions. This facilitates actionable insights and robust analytics, ultimately elevating retail experience.

Read our latest AI TechCompass to learn more.

Know in detail about trends in computer vision.

?