Self - Supervised Video Representation
Zhiyue Li
Is instance discrimination making the best use of data ( video )
Complementary view of video can be used to boost the gap between various visual - only video clip instances of the same class, which can be applied to generate positive training sets so that the general performance can be improved over InfoNCE instance training for video representation. Similarly, this idea can also be lead to application such as images and other modalities.
Page Views:
Is instance discrimination making the best use of data ( video, image) ? The answer is no. For the visual - only video of the same class, motion and RGB can provide complementary information about instance, which can increase the performance of positive instance. Both the RGB and flow motion can be improved, e.g. Generative Adversarial Networks (GANs).
