MobileGoturns: Light-Weight Deep Regression Networks for High-Speed Visual Feedback System
Chanrathnak BORANN, Hiromasa Oku · The Proceedings of JSME annual Conference on Robotics and Mechatronics (Robomec) · 2022
We present a light-weight CNN-based object tracker called MobileGoturn. The model resembles GOTURN except that convolutional layers are replaced by a low-cost structure called depthwise separable convolution and a hyper-parameter known as width-multiplier were applied to make a trade-off between speed and accuracy. We incorporated our novel model into a high-speed visual feedback system and demonstrated that the model is comparatively faster to track our proposed target objects in real-time.