Clothes Matching And Recommendation In Video Show

The E-commerce has been developing very fast in just ten years,and it breaks the space constraint that people have to go to the physical store to buy stuff. But now, when shopping online, people usually adopt the passive searching way, and initiative recommending is not mature yet.But initiative recommending commodity which is not equipped with text description is very significant for both merchants and consumers.This paper focused on clothes matching and recommendation in video show. Firstly, this paper used pedestrian detection and face detection to locate model, in which process the size threshold of regressed detection bounding box is adjusted. This paper built a Gauss mixture model to remove the background pixels and filtered out human skins through color space conversion. Secondly, this paper applied feature selection to manufactured features. This paper added color information to HOG feature in RGB and HSV ways respectively and improved the TOP1 matching correct rate of HOG. This paper applied BAG OF FEATURE algorithm to SIFT and SURF, in which process the best clustering number is tested, and improved the TOP1 matching correct rate both. Thirdly, this paper introduced deep learning to show video scene and extracted its feature, which has more semantic meaning. In this way,the TOP5 correct rate reached 99.1%. Then, to make the outcome robust and stable, this paper introduced person re-identification and invented a voting mechanism to improve the overall correct rate. Finally, this paper built a dataset, which contains 13 show videos and 9569 clothes photos,and trained a similarity network to make the TOP20 correct rate 22.8%.