Abstract: This blog post offers an overview of the 2019 paper “Objects as Points”, proposing a keypoint-based approach to object detection by modeling objects as bounding box centers with regressed attributes. It explains the principles of heatmap peaks for centers and direct regression for size/offset; the fully convolutional structure with multi-head outputs; implementation details like focal loss variants and backbones such as Hourglass/DLA; and experimental highlights on COCO, including superior speed-accuracy trade-offs and extensions to 3D detection and pose estimation.

Abstract: This blog post offers a exploration of the 2017 paper “Attention Is All You Need” by Ashish Vaswani and colleagues at Google, introducing the Transformer architecture that replaces recurrent networks with self-attention mechanisms for sequence transduction tasks like machine translation. It explains the core principles, including scaled dot-product and multi-head attention; the encoder-decoder structure with positional encodings, residual connections, and layer normalization; and implementation details such as training optimizations, emphasizing the model’s parallelizability and efficiency.

Abstract: Scene flow, which is a 3D version of optical flow, represents how each point in an image or point cloud changes in the two preceding and following frames. One paper in CVPR 2021 illustrates a weakly supervised learning of Rigid 3D Scene Flow approach. This article focuses on documenting running process of the official code.

Abstract: In order to improve the performance of deep learning inference, the deep learning inference acceleration technology has made remarkable progress in recent years. It effectively accelerates the speed of deep learning inference by optimizing the network structure, data, algorithm and hardware. This article will introduce several common deep learning acceleration techniques.

Abstract: The blog will introduce the whole process of how to install CUDA, TensorFlow, and MXNet environment on a Ubuntu computer. Demos of testing if the installations are successfull will also be showed. The edition of the software are CUDA 9.0, TensorFlow 1.5, MXNet-cu90 1.2 and Ubuntu 17.10.

Abstract: Parallel computing can speed up our programme. From data parallel to process parallel, there are many parallel tool and way that we can use. The blog will introduce The Parallel Patterns Library (PPL), Open Multi-Processing (OpenMP) and Open Computing Language (OpenCL). Other important parallel tools such as CUDA and Hadoop will be introduced as topics in later blogs.

Abstract: Decision tree algorithm is a very basic kind of algorithm in machine learning, and it is also an important tag learning method . This blog explains the basic principles of the decision tree algorithm and several algorithm variants. A simple example is also implemented to show how to construct a decision tree.

Abstract: The book Deep Learning(by Goodfellow, Bengio & Courville) has been read through. This blog is a reading note which record some key knowledge and points for memos.

Abstract: The book Dark Time(by Weipeng Liu) has been read through. It sparked me a lot. This blog will list some of the points in the book for not only sharing but also a reminder.

Abstract: This is a tutorial. This tutorial tells how to establish a personal blog based on github.io static pages with Hexo. The article also involves some setting details which can make your website more intimate.