how can I use the features of the middle of the model

[bmnptnt]

Hi, I'm trying to make a VCM(video coding for machine) sequence for object detection.
I use the faster_rcnn_X_101_32x8d_FPN_3x model for object detection.
But for VCM, I need features of middle of the model (after convolution).
And I also need to know How to put these features back into the middle of the model as input.
Because I have to use video coding between feature extraction and detection.
So, if you know a solution, please let me know.