[Feature] Incorporating TF-Blender for VOD

[khurramHashmi]

1. Incorporated TF-Blender into Video Object Detection methods.
2. The source code is borrowed from the original authors (HERE)
3. The module TF-Blender is crafted in the FGFA pipeline for now.

[GT9505]

Hi~@khurramHashmi , very appreciate your contribution to support TF-Blender method.

Please refer to contributing guides in order to fix the lint error.

Besides, we also have unittests to verify the correctness of added modules. It would be great if you could add unittests for the TFBlenderAggregator module.

[Chop1]

1. Incorporated TF-Blender into Video Object Detection methods.

   2. The source code is borrowed from the original authors (HERE)

   3. The module TF-Blender is crafted in the FGFA pipeline for now.

Hello, I have a question regarding the proposed implementation. I know it is borrowed from the original author.

In the description you write 1. Building an aggregated tensor from x, x_embed ,ref_x, and 'ref_x_embed' of shape [N, C*8, H, W] 2. Compute weights through passing Temporal Relation, Feature Adjustment, and Feature Blender modules. 3. Use the normlized (i.e. softmax) cos similarity to weightedly sum ref_x.

so the aggregated tensor correspond to tf_weight = torch.cat(...)

but the 2nd step I struggle to see it.

The temporal relation module correspond to what is named tf_blender

for tf_blender in self.tf_blenders:
            tf_weight = tf_blender(tf_weight)

then I fail to see the Feature Adjustment and Feature Blender modules The rest of the code seams to be the 3rd step.

Thanks in advance for your response

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}