deepmind
flow_imagenet checkpoint
The readme states:
The default model has been pre-trained on ImageNet and then Kinetics
As far as I understand, "pre-trained on ImageNet" means 3D convolutional NN which weights are recieved by bootsrapping values from 2D convolutional NN trained on ImageNet. So data/checkpoints/rgb_imagenet checkpoint is RGB network which is initialized with bootstrapped weights from 2D NN (which was trained on RGB images from ImageNet) and then trained on Kinetics dataset. Please, correct me if there are any mistakes.
If the description above is correct then I'm not sure I understand how data/checkpoints/flow_imagenet weights were achieved. I think it is Flow network which is initialized with bootstrapped weights from 2D NN (which was trained on RGB images from ImageNet) and then trained on optical flow values from Kinetics dataset. It's suprising that flow NN is initialized with RGB values. Is it really so?
