set params when training the flow model

[arunos728]

I'm trying to reproduce the two-stream results of TSM on Something v1, but the performance of my flow model is far below. (segment based sampling method)

I understand the 10 channels stacked optical flow (TV-L1) / learning rate 5 times in the first conv layer.

Is there any difference between RGB and Flow model in setting params?
(e.x, epochs, learning rate...)

[tonylins]

Hi, we basically used the same training hyper-parameters for flow models. How much accuracy can you get using optical flow? Note that optical flow model should give a lower performance compared to RGB input.

[arunos728]

I just stopped the training process before 50 epochs cuz my performance is too low.
It was lower than 30% at 40epochs.
Could you tell me the performance of TSM Flow model on Something v1? (uniform sampling, 8 segments, single inference, ResNet-50)

[tonylins]

The accuracy of TSM Flow model on Something V1 is about 42%.

[arunos728]

Thanks for the answer.
I just modified the dense sampling to the uniform sampling, and achieves the similar performance as you said.

[boheumd]

Thanks for the answer.
I just modified the dense sampling to the uniform sampling, and achieves the similar performance as you said.

Hi, I just trained the TSM-ResNet50 model on SomethingV1 with 8-segments optical flow input only, but the performance is around 37%, far below the 42%, and the two-stream is around 48%. Can you provide the link to download the optical flow for SomethingV1 dataset that you use?

[jianghaojun]

@boheumd How did you extract flow data for sthv1?