some question about backbone net input size?

[jcyhcs]

hi,professor:
i fllow your code, and try convert to tensorrt code, i convert backbone net to onnx format ok, i use netron , then input size is

,but i debug your ncnn-macos code, i find code below:
ncnn::Mat ncnn_img = ncnn::Mat::from_pixels(z_crop.data, ncnn::Mat::PIXEL_BGR2RGB, z_crop.cols, z_crop.rows);
printf("ncnn_img.w:%d,ncnn_img.h:%d\n",ncnn_img.w,ncnn_img.h);
the result is ncnn_img.w=127, ncnn_img.h=127;
so, what the real size for backbone input node??
how can i deal with this problem?
PLEASE HELP!

[HonglinChu]

hi,professor: i fllow your code, and try convert to tensorrt code, i convert backbone net to onnx format ok, i use netron , then input size is ,but i debug your ncnn-macos code, i find code below: ncnn::Mat ncnn_img = ncnn::Mat::from_pixels(z_crop.data, ncnn::Mat::PIXEL_BGR2RGB, z_crop.cols, z_crop.rows); printf("ncnn_img.w:%d,ncnn_img.h:%d\n",ncnn_img.w,ncnn_img.h); the result is ncnn_img.w=127, ncnn_img.h=127; so, what the real size for backbone input node?? how can i deal with this problem? PLEASE HELP!

NCNN supports adaptive scale input. And the initial template size is [1,3,127,127], search template size is [1,3,255,255]

[jcyhcs]

hi,professor:
i have success run NanoTrack on jetson tx2 with TensorRT accelerate! it very fast, fps almost 77fps~122fps!
but i test the dance_girl.mp4, i find the track not very good, always have lost and track error!
the result is same with ncnn version.
i upload the result video of dance_girl, please tell me why NanoTrack is poor performance, and how to make it more better?

dance_girl_result.mp4