Binarized Mode Seeking for Scalable Visual Pattern Discovery Wei - PowerPoint PPT Presentation

Binarized Mode Seeking for Scalable Visual Pattern Discovery Wei Zhang, Xiaochun Cao, Rui Wang, Yuanfang Guo, Zhineng Chen 张炜 中国科学院信息工程研究所 http://vireo.cs.cityu.edu.hk/wzhang

Background – Big Vis isual Data 3,320 images uploaded / minute 100 hours videos uploaded / minute 350 million new photos / day 1000 hours of TV News / day 2

Background – Search & Mining How to manage large volume of data? • Applications : browsing / recommendation / visualization / tagging … • Key techniques: searching and mining search mining/discovery Dataset query frequent items relevant items … … 3

Visual Pattern Discovery Given an unsorted data collection (e.g., > 10 8 ) • What kind of images are in the dataset? • What’s the difference with other ‘common’ datasets?

Main Challenge difficulty small patterns, large collection easy hard Previous works data-scale • Object-level • Optimization - Tan09; Yuan07; Liu10 Zhang17 10 8 • FIM - Quack08 10 7 • large dataset clustering - Sivic04 Chum10 Zhang14/16 • hashing: Chum09 10 6 Philbin08 Chum09 • ToF: Zhang14/16 – small pattern 10 5 • Frame level Sivic04 10 4 • graph clustering: Philbin08 Quack08 • Liu10 hashing: Chum10 small dataset 10 3 Yuan07 10 2 Tan09 Zhang17 – large collection 10 1 object-level frame-level pattern-scale

Binarized Mode Seeking for Scalable Visual Pattern Discovery Zhang17 – large collection • • Informative patterns Scalability Issue • • Memory: 10^6, FC7 feature  30G Frequent • • Discriminative CPU bottleneck 10110100001011 10100011101010 00101100110111 bMS 10101101001010 00110110101010 10101101001010 00111111010101 00101010101000 10111000110010 ... ... 00111011101000 frequent patterns ... ... target set 10110100001011 cbMS 10101101011010 10100011101010 00101010101000 00101100110111 00101011101000 00110110101010 ... ... 00111111010101 00111011101000 ... ... informative patterns contrastive set (d) patterns (a) large image sets (b) binary codes (c) modes

bMS : Binarized Mode Seeking Binary constraint Binomial-based kernel

Results on ILSVRC and Flickr

cbMS : Contrastive Binarized Mode Seeking Contrastive density • Positive set: the target set • Negative set: adopted as reference

cbMS : Contrastive Binarized Mode Seeking Contrastive density • Positive set: the target set • Negative set: a reference set bMS

Evaluation ILSVRC: 1.3 million images 1000 objects • No significant performance drop • Runs much faster

Patterns discovered on ILSVRC

Example patterns on Flickr

Summary Binarized Mode Seeking for Scalable Visual Pattern Discovery • Given a dataset with TOO MANY images (can’t fit into memory) • What kind of images are in the dataset? • What’s the difference with other “common” datasets? • Binarized data / algorithm • 50X CPU speedup • 30X memory saving Comparison of bMS and cbMS in ILSVRC. Middle: sample images from Patterns auto-discovered by bMS (left) and cbMS (right) from the other categories that are visually similar to the left column. unlabeled Flickr10M dataset.

Google 将在北京招聘更多的人工智能、机器学习技术专家， 该团队将与 Google Cloud 首席科学家李飞飞紧密合作，致力 于人工智能、机器学习领域的研究和技术开发。如果您有以 下领域相关经验：计算机视觉、视频处理、图像处理、语音 识别、自然语言处理、机器学习、深度学习、 HCI 、数据挖 掘等，欢迎申请 Research Scientist 职位！这是 Google 首次在 华招聘科学家 , 机会难得！有任何问题，或者需要内推，可以 直接小窗口联系 Lu Jiang, 或者将简历发送到 lujiang@cs.cmu.edu. JD 链接在此：