Yu Feng · Mar 11, 2020
Kuantum Leap, a joint venture of VizSeek for the China business and one of Kuan Capital's portfolio and incubated startups, has developed a proprietary feature vector-based shape recognition algorithm. This algorithm possesses the capability to accurately identify and locate billions of objects across various input formats, including JPEG, PDF, OBJ, MOV, and more. The algorithm has gained widespread approval and adoption among manufacturers and industrial distributors.
Often, our attraction towards purchasing something is driven by visual appeal. We desire to find similar items based on what we see. Traditional text-based searches may not suffice in such cases. Instead, why not take a picture and search for the most visually similar item? This is made possible by visual search technologies, which we commonly encounter on platforms like Google and Bing.
The prevailing approach today involves utilizing AI to analyze shape, color, and size to deliver the most relevant search results. To achieve accurate results, vision AI needs to be trained, much like teaching a child. Researchers and engineers train vision AI engines by exposing them to numerous classified and labeled images. The AI engine processes each pixel of these images and learns from them, refining and expanding its understanding of different objects over time. However, this pixel-based and machine-learning approach does have its limitations. In untrained areas, the search results may be perplexing, especially when it comes to rare and purpose-built industrial or art objects.
Let’s breakdown the search process and walk through an example of ranking world cities by proximity to a city given as input. First, the feature vector is computed for each city in the database using its latitude and longitude. Then each feature vector is positioned in an N-Dimensional space (N = number of dimensions of the feature vector = 2) as shown in Figure 1. Now, let’s say we would like to know which city in the database is closest to Tokyo. We would compute the feature vector of the search input which is Tokyo, position Tokyo in the 2-Dimensional space, and then measure the distance between Tokyo and each city in the database. Beijing would be returned as the closest match, and earn the number one rank in the search results.
Figure 1: 2-Dimensional space used for searching world cities by proximity.
The same logic applies during Kuantum Leap visual search as shown in Figure 2. For each shape (image, 2D drawing, or 3D model) in the database, Kuantum Leap computes its feature vector, and positions each shape in an N-Dimensional space (N > 2). What differs between this example and the world cities example are the defined rules that are being used for the computation of the feature vectors. If the visual search input being uploaded is a square, Kuantum Leap computes the feature vector of the input square, and positions it in the N-Dimensional space. Kuantum Leap then measures the distance between the input square and each shape in the database. Kuantum Leap would then return the square as the most accurate match, and assign it the number one rank in the search results.