The Bear at Tolo Hill: a blog about computing machines and lifelong learning.

  • Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition

    Today we released the paper, together with code and modes, of our AAAI 2018 publication “Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition“. This work is one of my most liked papers in recent years. It provides an elegant way to deal with temporal dynamic graphs such as skeleton sequences. Without any feature or…

  • Kinetics Pretrained TSN Models Released

    Kinetics Human Action Video Dataset is a large-scale video action recognition dataset released by Google DeepMind. It contains around 300,000 trimmed human action videos from 400 action classes. This year (2017), it served in the ActivityNet challenge as the trimmed video classification track. During our participation in the challenge, we have confirmed that our TSN framework published in ECCV 2016 works smoothly on…

  • Squeezing Memory out of Caffe

    No, I am not talking about thinking of the good old days with a cup of drink. Today we are talking about our recent release of a new feature in our modified version of deep learning toolbox, Caffe. In this release, we have implemented the functionality usually referred to as “memory multiloading”. A Wiki…

  • ECCV Paper: Temporal Segment Networks: Towards Good Practices in Deep Action Recognition

    We just released the code for Temporal Segment Networks (TSN) accepted to ECCV 2016. Github: Temporal Segment Networks In this release, we not only want to showcase the research work, but also mean to provide an accessible framework that will lower the barrier of entrance for research in action recognition or more general video understanding.…

  • 自动维护的文章列表

    下面这个网站提供基于DBLP自动生成嵌入式文章列表的功能。这样在自己的主页上就可以嵌入一个不需要维护的文章列表了,还是很方便的。 我的大概长这样 Yuanjun’s publication

  • Caffe Install Guide Ubuntu 15.10

    For those who are seeking help for installing Caffe on Ubuntu 15.10, here comes the cure. Linked is a detailed guided tour from our good friend, Ruohui Wang. Compile and run Caffe on Ubuntu 15.10

  • Intel 终于也注意到deep learning的问题了

    Single Node Caffe Scoring and Training on Intel® Xeon E5-Series Processors With these optimizations time to train AlexNet* network on full ILSVRC-2012 dataset to 80% top5 accuracy reduces from 58 days to about 5 days. LOL.

  • Supercharge Caffe with MPI Parallelism

    Our fork of Caffe at Github Repo Features MPI-based data/hybrid parallelism, low communication overhead Almost transparent parallel support Scaling up to multiple GPU and machines, with Ethernet/Infiniband Video data input.

  • MPI Broadcast

    One interesting answer I saw on Stack Overflow. Someone asked how he can receive the msg send by `MPI_Bcast`. The answer is This is a common source of confusion for people new to MPI. You don’t use MPI_Recv() to receive data sent by a broadcast; you use MPI_Bcast(). For MPI collective communications, everyone has to particpate; everyone has to call…

  • Face Recognition新进展

    Facebook刚刚公布了他们在CVPR2014的Oral文章 DeepFace: Closing the Gap to Human-Level Performance in Face Verification 文章题目就取得很霸气。Facebook的人用四百万人脸图片训练了一个九层的卷积神经网络(CNN),在著名的公共测试数据集LFW(labeled face in the wild)上达到了97.25%的正确率。这个数字已经基本接近人眼的辨识水平。