Self-Training and Multi-Task Learning for Limited Data: Evaluation Study on Object Detection

  • 2023-09-12 15:50:14
  • Hoàng-Ân Lê, Minh-Tan Pham
  • 0

Abstract

Self-training allows a network to learn from the predictions of a morecomplicated model, thus often requires well-trained teacher models and mixtureof teacher-student data while multi-task learning jointly optimizes differenttargets to learn salient interrelationship and requires multi-task annotationsfor each training example. These frameworks, despite being particularly datademanding have potentials for data exploitation if such assumptions can berelaxed. In this paper, we compare self-training object detection under thedeficiency of teacher training data where students are trained on unseenexamples by the teacher, and multi-task learning with partially annotated data,i.e. single-task annotation per training example. Both scenarios have their ownlimitation but potentially helpful with limited annotated data. Experimentalresults show the improvement of performance when using a weak teacher withunseen data for training a multi-task student. Despite the limited setup webelieve the experimental results show the potential of multi-task knowledgedistillation and self-training, which could be beneficial for future study.Source code is at https://lhoangan.github.io/multas.

 

Quick Read (beta)

loading the full paper ...