Parameters and variables setting in DFF model

[yan811]

In "/mmtracking/mmtrack/models/motion/flownet_simple.py," the init parameters "flow_img_norm_std=[255.0, 255.0, 255.0]" and "flow_img_norm_mean=[0.411, 0.432, 0.450]" . What's the meaning of these parameters? I'm using a type of data with 10 channels, how should I set these parameters?

Also, in "prepare_imgs" method, "img_metas[0]['img_norm_cfg']['mean']" and "img_metas[0]['img_norm_cfg']['std']" are both initialized with 0.Is it necessary to reassign the value while training or testing? If necessary, how and what value should I assign to these variables?

[dyhBUPT]

Hi, the std and mean are used to nromalize the input img.

About how to set these parameters, it depends on what the 10 channels of your data are.

[yan811]

I got it. Is there any API to calculate the std and mean of my dataset (cocovid format)?

[dyhBUPT]

I'm sorry but I don't know about it.

[yan811]

All right. Thanks.

[yan811]

I am not clear yet. In flownet_simple.py, there are two types of std/mean: flow_img_norm_std/flow_img_norm_mean; img_metas['img_norm_cfg']['mean']/img_metas['img_norm_cfg']['std']. I think the latter ones refer to the input dataset. But I'm still wondering what the previous ones refer to.

[dyhBUPT]

Hi, the flownet has its own mean and std for normalization, and the effect of img_nrom_cfg is cancelled out. For more details, you can read these codes: https://github.com/open-mmlab/mmtracking/blob/bcc0c9b3a3062c770457989e240cb55c3b667875/mmtrack/models/motion/flownet_simple.py#L183-L191 I hope this will help you.

Best wishes.

[yan811]

Normalization is usually calculated as: "(img-img_mean)/img_std", but in line 183-184, the data is transfered with "img/std-mean", which is not the same form as normalization. Is there any other meaning of this transformation?

Also in these codes, self.flow_img_norm_std=[255.0, 255.0, 255.0], self.flow_img_norm_mean=[0.411, 0.432, 0.450]. They both have 3 channels. But my 10-channel data requires 10-channel flow_img_norm_std and 10-channel flow_img_norm_mean. How to set these params?

[dyhBUPT]

line 183 img * std + mean is used to cancelled out the effect of (img-mean)/std. line 184 img / std - mean is ok, because the mean is in [0, 1], not [0, 255]. In summary, there is a simple mathematical manipulation.

For your 10-channel data, maybe you can calculate it's statistics in each channel (i.e., mean and std) and set them as the nrom params.

[yan811]

I think my data's statistics are supposed to set the self.img_norm_std and self.img_norm_mean. But as mentioned before, flow_img_norm_std and flow_img_normmean are the statistics of flownet. How to set the flow..._params to 10-channel?

[dyhBUPT]

I give an example, for one channel, I suppose that mean=100, std=5. Then you can set as:

img_nrom_mean = 100
img_nrom_std = 5
flow_img_norm_mean = 20  # i.e., 100 / 5
flow_img_norm_std = 5

Not that I'm not sure if this works. You can study it in detail.

[yan811]

In your example, img_norm_mean=mean,img_norm_std=std, flow_img_norm_std=img_norm_std=std, flow_img_norm_mean=img_norm_mean/img_norm_std Is that right?

[dyhBUPT]

Yes, but please refer with caution. because I'm not so confident for your case.

[yan811]

I will try it. Thanks.