wangdomg Goto Github PK

social_lstm_pedestrian_prediction

The aim of the project is to predict the trajectories of pedestrians using lstm neural networks. The project starts from the paper "Social LSTM: Human Trajectory Prediction in Crowded Spaces - Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, Silvio Savarese - Stanford University - The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 961-971", and its official implementation (https://github.com/vvanirudh/social-lstm-tf) and makes some modifications. This is the Multimedia communication course project made during my final year of the bachelor degree under the supervision of professor Nicola Conci and his phd student Niccolò Bisagno. The modifications introduced are two: - To every simulated pedestrian add the input goal; the goal is the final position (in x and y coordinates) of that pedestrian when it disappears from the video. This modification should improve the predicted trajectory of that pedestrian because of the introduction of this new information - The grid created for every pedestrian in the original project to identify nearby pedestrians is replaced with an array containing the position(in x and y coordinates) of the others pedestrians in distance order, from the closest to the farther. This modification should improve the model results beacuse it presents relevant informations in order to the neural network. Then these two modifications were combined in to a single model. Every model has been evaluated in the test videos with different parameters and in conclusion the model with the two modifications (goal and array) combined performed better than any other model. Also the two modify models performed better than the original model. These results can be seen in the report at page 12 and 13. Unfortunately I haven't the time to translate the report in english, because now is in italian, but the result table at page 12 an 13 should be pretty clear. Technical details: - Programming language: Python 2.7 - Neural networks library used: Tensorflow 1.5 - External libraries: CUDA 8.0, CUDNN 6.0 - OS: Linux, Ubuntu 16.04 distrubution License: GPL v3

speech-demo

语音api示例

stanford-nlp-usage

整理了Stanford NLP的部分使用方法

strnn

Predicting the Next Location: A Recurrent Model with Spatial and Temporal Contexts

style-transfer-in-text

Paper List for Style Transfer in Text

suggested-notation-for-machine-learning

This introduces a suggestion of mathematical notation protocol for machine learning.

torchsample

High-Level Training, Data Augmentation, and Utilities for Pytorch

tpu

Reference models and tools for Cloud TPUs.

transformation

userprofile

今日头条的一个比赛，做文章分类

uu_cf

Python实现的基于User的协同过滤算法

web_kg

爬取百度百科中文页面，抽取三元组信息，构建中文知识图谱

wide-and-deep-pytorch

PyTorch implementation of Tensorflow's Wide and Deep Algorithm

word2vec

用C++实现的Word2Vec

word_similarity_and_word_analogy

Word Similarity and Word Analogy Task scripts

xflow

Distributed LR、 FM and MVM model on Parameter Server. FTRL and SGD Optimization Algorithm.