Temporal Activity Detection in Untrimmed Videos with Recurrent Neural Networks
2017-12-05 19:55
417 查看
团队介绍
作者：
Alberto Montes, Amaia Salvador, Santiago Pascual, Xavier Giro-i-Nieto作者都来自Universitat Politècnica de Catalunya (UPC)的一所西班牙大学，理工科挺强的，发表在NIPS workshop的一篇文章. 在ActivityNet Challenge 2016取得了不错的
动机
利用C3D[1]能捕捉短时间的空时特征，然后LSTM处理长时间的信息，Untrimmed Videos对进行分类和定位框架
C3D首先在sports-1M上进行预训练，之后对预处理好的每16帧视频（相邻的视频片段没有交叉）离线提取时空特征，将这些固定的特征作为LSTM的输入，进行每一个片段的分类，每一类对应于一个动作类别（增加background作为一类）。作者在文中也探讨了不同深度，不同宽度的LSTM网络。发现1x512的最浅最窄的最好。The simplest is the best.
tricks
对LSTM输出的概率进行均值滤波，使其更平滑，消除异常概率值为了应对背景数据较多的情况，在计算loss时候，给其相对较小的权重
思考改进
如果C3D和LSTM一起训练，微调C3D，重新训练LSTM，效果应该会好一点，但用于参数众多，也可能导致参数过多，导致过拟合引用
D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri, Learning Spatiotemporal Features with 3D Convolutional Networks, ICCV 2015
内容来自用户分享和网络整理，不保证内容的准确性，如有侵权内容，可联系管理员处理 
相关文章推荐
- 论文《Inside-Outside Net: Detecting Objects in Context with skip pooling and Recurrent Neural Networks》
- Mitosis Detection in Breast Cancer Histology Images with Deep Neural Networks
- 极简笔记 DeepID-Net: Object Detection with Deformable Part Based Convolutional Neural Networks
- 论文阅读：Reading Text in the Wild with Convolutional Neural Networks
- Towards End-to-End Car License Plates Detection and Recognition with Deep Neural Networks
- [行为识别论文详解]SSN(Temporal Action Detection with Structured Segment Networks)
- 车牌识别“Towards End-to-End Car License Plates Detection and Recognition with Deep Neural Networks”
- READING NOTE: Object Detection from Video Tubelets with Convolutional Neural Networks
- 【每周一文】Supervised Sequence Labelling with Recurrent Neural Networks
- Chinese Poetry Generation with Recurrent Neural Networks
- 多尺度R-CNN(2): Inside-Outside Net: Detecting Objects in Context with Skip Pooling and Recurrent Neural
- Scene Labeling with LSTM Recurrent Neural Networks
- Temporal Action Detection with Structured Segment Networks (ssn)【转】
- 车牌检测识别--Towards End-to-End Car License Plates Detection and Recognition with Deep Neural Networks
- 【论文笔记2】图像压缩神经网络在Kodak数据集上首次超越JPEG——Full Resolution Image Compression with Recurrent Neural Networks
- 车辆计数“FCN-rLSTM: Deep Spatio-Temporal Neural Networks for Vehicle Counting in City Cameras”
- SSN：Temporal Action Detection with Structured Segment Networks
- (zhuan) Attention in Long Short-Term Memory Recurrent Neural Networks
- 论文阅读：Reading Text in the Wild with Convolutional Neural Networks
- Object Detection from Video Tubelets with Convolutional Neural Networks