Since then the retail sector in recent years to upgrade constructively put forward a new blueprint with three retail core "of people, goods, field" based around, the parties of new technologies in the retail rush wrestling each type of scene. Before sharing, we introduced the singular point cloud is how artificial intelligence technology, from the "human" dimension cut into the retail scene. The article will focus on artificial intelligence technology singularity is how cloud enabling retail from "goods" business dimension.
The first step in visual intelligence as a singular point cloud AI engine landing, plays a major role in the digitization of commodities. From a visual perspective, the "goods" morphological differences compared to the "people" who have been significantly increased. Different vertical industries to adapt to market demand the development of products, both in product features or forms are vastly different, even the same type of product will produce different forms in order to comply with differentiated demand segments. This presents new challenges for visual intelligence engine, but also stimulate the StartDT AI Lab greater technological breakthroughs. The following selected a few more representative of the industry has been a scene from a singular point cloud commercial landing scenarios to demonstrate technological breakthroughs StartDT AI Lab acquired.
AI know what energized the apparel industry
We pioneered garment identification technology in 17 years, combined with the product - singularity Mirror achieved floor. After this scenario, the user standing in front singularity mirror stood still for a few seconds, the user will first singular point of the mirror divided wearing understood isolated from various types of apparel T-shirts, coats, jackets, jeans and then through the analysis of single items of clothing, such as T-shirts, we can extract its characteristics, analyzed T-shirt style, sleeve length, version and other attributes, the last in our self-built one million apparel items in the library, use the recommended algorithm recommended for users of a similar, matching merchandise, so as to achieve the effect of drainage client, the smart shopping guide.
In this link, the main problem we had encountered and solutions:
✨1. When acquiring the user information and to obtain accurate characteristics of clothing, like T-shirts, shorts garment so obvious upside down more readily available, but like dresses, long coat class clothing, but difficult to handle.
We collect a lot of data, marking clean, improve data collection and use cascade method, the first of a clothing label were detected and analyzed, and then a second treatment, thereby increasing segmentation.
✨2. Clothing attribute diverse and there is no uniform standard, while the judge how quickly become a problem.
We attribute are independent of each classifier, a feature of the input repetitive operations is reduced, In addition, we will classify simultaneously connected to the Graph, the entire process of End-to-End, quickly and efficiently.
. ✨3 large-scale retrieval: when the database is large, the retrieval speed is slow, can not respond quickly.
Our database than to be deployed on a distributed cluster, to achieve a feature matching level map-reduce, so that when we deal with various levels of ease of comparison.
AI know what energized FMCG beverage industry
We explore the history of nearly two years in FMCG beverage industry. Wide range of drinks - including mineral water, carbonated soft drinks, juices, beer, liquor and so can be sold in Singularity magic cabinet. Through self-development of deep learning algorithms, data sampling method, with our customized hardware configuration to achieve the recognition accuracy rate of more than 99% in the commercial scene. We'll cargo damage control in less than 1%, lower than the industry average FMCG drinks cargo damage level. In addition, we are on less than one square meter of floor space to achieve a very high proportion of Ping, a single cabinet single-thousand-level retail sales up, but only a small amount of maintenance work.
Behind the official business, StartDT AI Lab main contribution are the following:
No.1 fast and efficient algorithm to detect small objects
And unlike most scenarios, our goal is often detected near the edge of the smallest of small targets 16 pixels, and there are quite a few goals on a graph. We Adaptive anchor, the anchor so that more accurate a priori; In addition, we one characteristic enhancement algorithm, characterized in that the details of the network to reduce losses in the depth as far as possible, by avoiding too few features brought low score the detection result of instability. We also study from a self-distillation method, without increasing parameters to enhance the accuracy of the model, so as to achieve the commercial level.
No.2 recognition technology with metric Learning
Measure learning is widely used in the past few years in the face recognition model, and achieved very good results. We incorporated merchandise recognition, binding of classic neural network, so that more accurate and reliable the recognition result; output model may also comparable in addition to features, the results support the feature comparison mode, the comparison between different classes like support, so that the level of product selected from avoiding the sale of similar goods at the same time the problem is not recognized.
No.3 small data sets enhanced
Our data set is relatively usage scenarios, in fact, is a small data set. How to use a small set of data obtained commercial-grade accuracy in the big data scene? We mix a self-study method, the detection model to obtain a very high recall rate; in addition, we also use the GAN, in the process of training the classifier generator while training, a chance to produce results while training a classifier, the classification training is fuller, smoother.
In the vertical field of fresh / pharmaceuticals, we dare to explore and try to take advantage of deep self-learning algorithm and sampling methods, combined with the advantages of self-development of hardware, the first can complicated multi-class category is not affected by fresh intelligence containers. It is well known in the field of vertical, the same change in the appearance of diverse sku, high adaptation costs. Through specific product design, the perfect support for all kinds of fresh fruits and vegetables, more accurate rate can be done close to 100%. Customers are not limited to the specific constraints of the site markets / pharmacy, etc. At the same time, but no one can play the advantages of perfect container - without time limit.
(From top to bottom to identify pharmaceutical products, fresh class identification)
StartDT AI Lab to explore in the dark harvest and innovation:
No.1 deep learning style novel thinking scene
In order to get rid of under fresh scene, bringing the similarities and differences of various shapes and sizes and difficult compatible, as well as new customers quickly and effectively on the needs of us on the actual road scene landing, combined with the characteristics of goods, product superiority and depth of learning algorithms principle, to the grotesque of fresh products for special packaging, the solution to the differences between fresh products, and can support customers quickly on the new.
No.2 innovative features unique augmented data
There are light novel design enough to make a perfect landing deep learning in a real-world scenario, the customer is the first element of accuracy, data is depth study of capital. StartDT AI Lab after in-depth analysis of the data characteristics and performance of the algorithm vertical field of medicines and fresh, innovative way in augmenting the sampling data on a data set, allowing near-perfect accuracy, interpretation of what is to be landed deep learning the concept of.
From the above case, it presents a singular point cloud explore in the "goods" are digital, although they are still relatively preliminary stage of digital goods, but also by AI technology for the first time to achieve a full digital tracking link in the sale of commodities cycle. And on this basis to achieve front-end retail cost reduction, improved user research automation efficiency. In order to make business smarter, StartDT AI Lab will move on, please sustained attention ~