search

researchmm / LightTrack

[CVPR21] LightTrack: Finding Lightweight Neural Network for Object Tracking via One-Shot Architecture Search
MIT License
396 stars 59 forks source link

移动端平台测试相关 #7

Closed wwn1233 closed 3 years ago

wwn1233 commented 3 years ago

很不错的工作，支持； 另外想请教一下 在安卓平台部署的代码使用什么呢 可以分享学习实测一下吗？

MasterBin-IIAU commented 3 years ago

Thank you for your appreciation to our work :) We tried two frameworks: TNN and onnxruntime. We use the former one to deploy our models on the mobile phones. But the latter one supports more operations than the former one.

wwn1233 commented 3 years ago

our

good。 But how to deploy the correlation operation in your model? Is it a custom op implemented by yourself?

MasterBin-IIAU commented 3 years ago

Hi, sorry for the late reply. In LightTrack, we use "pixel-wise correlation" to aggregate features of the template and the search region. We provide two types of implementations, one is by "convolution and for-loop" and the other is by "matrix multiplication". The second one runs much faster than the first one because the computation is quite vectorized. But some frameworks may not support the "matrix multiplication" operation. In this case, you can turn to use the first one (conv+for-loop).

Thanks, Bin

wwn1233 commented 3 years ago

thanks！