视频分类模型整理

视频特征

Handcrafted Feature

• HOG(Histogram of Grident)
• HOF(Histogram of Feature)
• BoF(Bag of Feature)
• Dense Point Trajectory
• Motion Boundary History(MBH)
• Trajectory Fisher Vector

Spatial-Temporal Feature

• ConvNet Feature
• Stacked OpticalFlow

Audio Feature

• MFCC

视频分类方法

Trajectory-Based

• Action Recognition by Dense Trajectories. [Paper] [Code]
• Dense trajectories and motion boundary descriptors for action recognition. [Paper] [Python Code]

CNN-based

• Large-scale Video Classification with Convolutional Neural Network. [Project]
• Learning SpatialTemporal Feautres with 3D Convolutional Networks. [Paper]
• 3D Convolutional Neural Networks for Human Action Recognition. [Paper]
• Two-Stream Convolutional Networks for Action Recognition in Videos. [Paper][Code]
• Two-Stream SR-CNNs for Action Recognition in Videos. [Paper][Code]
• Trajectory-Pooled Deep Convolutional Descriptor. [Paper]
• Temporal Segment Networks: Towards Good Practices for Deep Action Recognition. [Paper]
• Learnable pooling with Context Gating for video classification.

动作检测

• Finding Action Tubes. [Paper] [Code]
• Multi-region Two-Stream R-CNN for Action Detection.
• Temporal Action Detection with Structured Segment Networks.
• UntrimmedNets for Weakly Supervised Action Recognition and Detection
• R-CNNs for Pose Estimation and Action Detection.

特征融合

• Multi-Stream Multi-Class Fusion of Deep Networks for Video Classification. [Paper]
• Real-time Action Recognition with Enhanced Motion Vector CNNs.[Project]
• Convolutional Two-Stream Network Fusion for Video Action Recognition. [Paper][Code]
• ActionVLAD: Learning spatio-temporal aggregation for action classification. [Project]