Andrzej Pronobis, PhD, Robotics and Machine Learning Researcher, Climber
Répondu il y a 297w
I would also like to point to a product called Productify which we recently developed in our company OculusAi. Productify is a mobile app that can recognize apparel products and recommend stores and other visually similar products. The goal is to focus on products that are visual in nature. Books for example have titles that are visible on the cover and make identification easy even without using a computer vision algorithm. The idea of Productify is to recognize products for which the only description is appearance e.g. shoes, bags, shirts etc.
Chomba Bupe, builds computer vision systems
Répondu il y a 134w · L'auteur dispose de réponses 1k et de vues de réponses 2.8m
Having gone through the answers already here, it seems there are great projects being developed concerning visual recognition of items such as clothing, shoes e.t.c. But I would like to discuss on the possible methods of achieving such recognition systems.
If what you are trying to make is image understanding then I have bad news, no computer vision system actually understands images, they just map the pixel values to class labels or something else. To achieve that kind of image understanding would require something that processes images as humans do.
A simple method for recognition of clothing would be a bag-of-words method, this is because clothes are highly textured and have very rich colors, this means using a bag-of-words model with key point detection algorithms for computer vision can work but cannot produce state-of-the-art results. It can be difficult to describe an image using this technique but may work at least for similarity search.
Thus think of automated image captioning techniques:
The caption can be, "a bottle shaped white dress with black stripes". To find a dress similar to the query, one just needs to look up clothes with the same or similar captions and rank them according to the best matches this can be achieved by a search engine. The bottleneck here is thus the automated captioning algorithms.
The current trend for automated image captioning is to feed features learnt using a ConvNet into a a recurrent neural net. I think the best explanation is from Google research blog and it's worth looking at. So in your case you can approach the visual recognition of clothing as:
The items returned in the final stage maybe similar items being offered. You can end at the captioning stage if you would only like a description of the item in question. As can be seen above the arrow [math]->[/math] represents a purely feedforward process, the improvements would be to introduce attention based feedback loops, to allow for a more flexible feature integration scheme. Such a flexible feature integration process would help improve the captioning stage and the rest of the recognition process.
Obviously this is a very hard task to achieve accurately by current computer vision and machine learning approaches. One thing about human vision is that features such as shape and color are processed separately, multiple features are then integrated through a visual attention process, this recognition scheme is very powerful and flexible.
Hope this helps.
Sudhir Kumar Singh, Founder at StileEye
Répondu il y a 290w
We have solved this problem at StileEye (StileEye | Fashion).
StileEye is a visual fashion discovery platform that understands any image in terms of intrinsic visual cues such as color, pattern, style, shape, and object. StileEye app allows users to window shop the Web: users can snap or upload a photo of an intriguing style and discover similar trendsetting outfits from their favorite celebrities, designers, retailers, and social network. Users can also design their own outfits and discover what is available out there.
The fashion industry presents an interesting challenge – how to make computers see fashion on the Web exactly as consumers see it online and in everyday life? To do this, one has to isolate and identify clothing and accessories within visually complex images like personal photos and magazine spreads. Traditional fashion apps rely on text descriptions and other forms of human intervention, which limits their scope. But we wanted to understand millions of fashion items on the Web. This required a powerful and fundamentally different set of computer vision algorithms. So, to achieve this we built a new kind of science that demonstrates how we can learn visually from web-scale data and I am the primary inventor of that.
Please checkout the website and mobile apps as per the links below:
Site internet : StileEye | Fashion
iPhone/iPad app: StileEye
Google Play Store: Applications Android sur Google Play
StileEye intro video:
Jayant Kumar, Wrote PhD thesis in Computer Vision
Répondu il y a 297w
And-Or graph representation as a matching model:
Composite Templates for Cloth Modeling and Sketching (CVPR 2006)
L. Zhao, ”Dressed Human Modeling, Detection, and Parts Localization”, Ph.D Dissertation, CMU, July, 2001.
Related useful work:
Computer Graphics Techniques for Modeling Cloth, IEEE Computer Graphics & Applications , pp. 28-42, 1996.
Clothing Cosegmentation for Recognizing People(CVPR 2008)
Deepti Ghadiyaram, Speaks the language of pixels.
Répondu il y a 298w
Here is a Fashion discovery engine built around the same idea - StileEye | Fashion
Joy Tang, travaille chez High-Frequency Trading
Répondu il y a 82w
Check-out Accueil Markable’s fashion recognition accuracy is No.1 in the world, doubled accuracy from the No.2