TensorRT Plugin data access

I’m currently working with a ssd-type caffe network and trying to optimize dataflow (w/o modifying network structure) using custom layer plugins. I’m wondering - is there any possibility to get data pointer inside the layer plugin for network input image, so in detection_output plugin i could return detected objects with an images, cropped from original image (along with coordinates and confidences)?

TensorRT Version:
GPU Type: RTX 2070
CUDA Version: 10.2
CUDNN Version:
Thank you!

Please refer to below links related custom plugin implementation and sample: