C++ implementation of papers Mobile Vision Transformer-based Visual Object Tracking (BMVC2023) and Separable Self and Mixed Attention Transformers for Efficient Object Tracking (WACV2024)

The original Pytorch implementations is available in these links: MVT and SMAT.

Why C++ implementation?

• Improve the model inference speed (fps)
• Single vs Multi-thread throughput analysis
• Lower memory footprint compared to Python
• Subsequent deployment on Andorid devices using JNI

Supported Frameworks

• MNN
• ONNX-Runtime

To-do

• TF-Lite
• NCNN