Daily News

From Home Furnishing Business

Embodied Intelligence Unveiled for Interior Design at CVPR

July 3, 2024  by HFBusiness Staff in Business Strategy, Designer Weekly, Industry   

21 in Seattle, brought together top researchers, scientists, and industry leaders to explore breakthroughs in computer vision, artificial intelligence, machine learning, deep learning, and much more.

Embodied intelligence emerged as a focal point at CVPR this year, representing an advanced field that merges computer vision, machine learning, and robotics. The demand for large-scale and high-quality datasets has become increasingly evident, whether for training humanoid robots or constructing robust AI models.

Coohom Cloud was established by Koolab Innovation Laboratory of MANYCORE TECH INC. in 2017, after realizing that many industries were on the verge of significant investments in high-quality data. MANYCORE TECH INC. is a leading global cloud based interior design software platform, Coohom Cloud leverages its parent company MANYCORE TECH INC.'s extensive indoor data resources, combined with high-performance rendering engines and advanced data processing technologies, to deliver realistic and physically accurate 2D and 3D interior datasets and services to the AI industry.

In 2018, MANYCORE TECH INC. collaborated with Imperial College London and the University of Southern California to introduce the InteriorNet, which was then the largest publicly available indoor scene dataset. Building upon this, the MANYCORE TECH INC.'s platform now generate over 400,000 3D designs daily and houses approximately 360 million 3D models, encompassing furniture, appliances, household items, and more. This forms a robust data foundation provided by Coohom Cloud for research in indoor environment understanding, 3D reconstruction, robotics simulation and beyond.

Regarding 2D image rendering technology, Coohom Cloud employs a proprietary rendering engine to capture image data in diverse indoor settings through adjusting camera parameters, path trajectories, lighting conditions, etc., resulting in 2D datasets in formats such as RGB, depth, semantic, normal, point cloud, and others. This capability allows Coohom Cloud to produce 300,000 sets of 2D datasets daily, serving as ample training materials for AI agents' navigation, visual perception, and environmental understanding capabilities.

To provide cost-effective data service solutions, Coohom Cloud utilizes its self-developed data processing engine designed for efficient data transformation, converting the accumulated database into formats suitable for AI training. This data engine can offer more comprehensive support on the physical, diversity, and data labeling fronts.

Physical Enhancement: Embodied intelligence's real-world interactive capabilities are a key indicator of its level of intelligence. Coohom Cloud integrates cutting-edge simulation platforms like NVIDIA Isaac Sim to create 3D environments with realistic physical properties, allowing robots to undergo extensive interaction testing within a virtual realm. This significantly reduces training costs, enhances training safety, and improves repeatability.

Environment Augmentation: Coohom Cloud's environment enhancement tool enriches the complexity of virtual environments through model element diversification, material variations, as well as lighting adjustments, etc. to simulate the diversity and complexity of the real world, enhancing the adaptability and generalization ability of robots.

Segmentation & Annotation: Coohom Cloud utilizes advanced synthetic data generation techniques to customize segmentation and annotation data, supporting various labels such as semantics, materials, spatial states. This approach significantly enhances data processing efficiency and accuracy.

Moreover, Coohom Cloud remains committed to exploring additional possibilities and value for computer vision and the entire AI field through technological innovation and industry partnerships.

Photo: Photorealism Comparison: Real World Image vs. Synthetic Image