search

rohitgirdhar / ActionVLAD

ActionVLAD for video action classification (CVPR 2017)
https://rohitgirdhar.github.io/ActionVLAD/
Other
216 stars 61 forks source link

I use matconvnet to reproduce your work,and I have some questions~ #15

Closed icyzhang0923 closed 6 years ago

icyzhang0923 commented 6 years ago

I have some questions about the action vlad:

  1. How to sample the 25 frames from each video, random or uniform?
  2. How to preprocess the frames in Tensorflow?
  3. How to combine the two stream? use SVM classifier?
  4. I'm not sure the clip gradients at 5 l2 norm, to calculate which gradients' l2 norm，accumalated gradients,or the average gradient over the minibatch?
rohitgirdhar commented 6 years ago
  1. We found nearly similar performance either way
  2. Please refer to preprocessing/vgg_ucf_preprocessing.py for details
  3. We used weighted averaging of the softmax probabilities (similar to previous methods)
  4. This is how we use the clip_gradients