xdxuefei/tdd

Trajectory-pooled Deep-Convolutional Descriptors

Trajectory-Pooled Deep-Convolutional Descriptors

Here we provide the code for the extraction of Trajectory-Pooled Deep-Convolutional Descriptors (TDD), from the following paper:

Action Recognition with Trajectory-Pooled Deep-Convolutional Descriptors
Limin Wang, Yu Qiao, and Xiaou Tang, in CVPR, 2015

Updates

• Dec 24, 2015
  • Release the second version of TDD (branch: cudnn2.0) compatible with latest caffe toolbox. Due to speedup brought by cudnn2.0 or above, TDD extraction is becoming more efficient.
• Jul 21, 2015
  • Release the first version TDD (branch: master) compatible with an older version of caffe toolbox.

Two-stream CNN models trained on the UCF101 dataset

First, we provide our trained two-stream CNN models on the split1 of UCF101 dataset, which achieve the recognition accuracy of 84.7%

"Spatial net model (v1)"
"Spatial net prototxt (v1)"
"Temporal net model (v1)"
"Temporal net prototxt (v1)"

TDD demo code

Here, a matlab demo code for TDD extraction is provided.

• Step 1: Improved Trajectory Extraction
  You need download our modified iDT feature code and compile it by yourself. Improved Trajectories
• Step 2: TVL1 Optical Flow Extraction
  You need download our dense flow code and compile it by yourself. Dense Flow
• Step 3: Matcaffe
  You need download the public caffe toolbox. Our TDD code is compatatible with the latest version of parallel caffe toolbox.
  Note that you need to download the models in the new proto format:
  "Spatial net model (v2)" "Temporal net model (v2)"
  
• Step 4: TDD Extraction
  Now you can run the matlab file "script_demo.m" to extract TDD features.

Questions

Contact

• Limin Wang