What is temporal jittering & temporal scaling?

[adityarp9]

Hello @dutran

I have been reading & trying to understand the following paper on gesture recognition: [(http://research.nvidia.com/sites/default/files/pubs/2016-06_Online-Detection-and/NVIDIA_R3DCNN_cvpr2016.pdf)]

Here, they mention they have used augmentations like:

temporal scaling (±20%), and jittering (±3 frames).

This repo also has an implementation for temporal jitter @

• https://github.com/facebook/C3D/blob/36c3f3156cdea7a62b8f81614f61a10ea64da8ea/C3D-v1.1/src/caffe/util/image_io.cpp#L132

This has got me confused about what exactly is frame jittering:

1. Is it equivalent to frame dropping at a specified sampling rate?
2. Is it increasing the number of frames by copying certain (specified or random) frames?
3. Is it changing the order of certain (specified or random) frames i.e. the same number of frames but a disturbed sequence of frames?

Doubts with temporal scaling:

1. Drop frames at random indices across the temporal dimension?
2. Drop frames uniformly with a fixed sampling rate?

Apart from the implementation mentioned above, I could not find any documentation that explains these two augmentations. Please help.

Thanks in advance.