Repository logo
 
Loading...
Profile Picture

Search Results

Now showing 1 - 1 of 1
  • A zero­shot learning method for recognizing objects using low­power devices
    Publication . Patrício, Cristiano Pires; Neves, João Carlos Raposo; Proença, Hugo Pedro Martins Carriço
    Zero­Shot Learning (ZSL) has been a subject of increasing interest due to its revolutionary paradigm that simulates human behavior in recognizing objects that have never seen before. The ZSL models must be capable of recognizing classes that do not appear during training, using only the provided textual descriptions of the unseen classes as an aid. Despite the vast benchmarking around the ZSL paradigm, few works have assessed the computational performance of the developed strategy regarding inference time. Furthermore, no work has evaluated the effects of using different CNN architectures, such as lightweight architectures, apart from the de facto standard ResNet101 architecture, and the feasibility of deploying zero­shot learning approaches in a real­world scenario, particularly when using low­power devices. Consequently, in this dissertation, we carried out an extensive benchmarking toward analyzing the impact of using lightweight CNN architectures on ZSL performance, allowing us to perceive how the ZSL methods perform in real­world scenarios, mainly when run in low­power devices. Our experimental results demonstrate that the impact on the ZSL accuracy is not significant when a lightweight architecture is adopted, indicating the effectiveness of such low­power devices in performing ZSL methods.