SVM with R(2+1)D Features

[murilovarges]

Hi @dutran,

I'm trying to reproduce the results reported in "A Closer Look at Spatiotemporal Convolutions for Action Recognition" and a doubt has arisen. Why don't you use SVM with features extracted from R(2+1)D net?

In the previous papers, you use (Features from net + SVM): C3D - VGG Architecture - Learning Spatiotemporal Features with 3D Convolutional Networks Res3D - ResNet Architecture - ConvNet Architecture Search for Spatiotemporal Feature Learning

I'm trying to apply these architectures to detect pornography in videos and would like to know if is best use Features extracted from CNN + SVM or CNN + Softmax?

[dutran]

1. We did not try R(2+1)D features with SVM. For the earlier papers, the main focus is feature learning, thus experiments with linear SVM seems natural. Our R(2+1)D paper is more focusing on exploring architectures, thus fine-tuning is more relevant.
2. For your problem, fine-tuning will give you the best performance (in my opinion).