Scepter914 Website
Table of contents

RegTTA3D: Better Regression Makes Better Test-time Adaptive 3D Object Detection (ECCV2024)

Summary

  • Test-time adaptationを用いた3D detection Domain adaptation
    • Regressionを中心に少ないparameterでtuningできる

Method

  • Domain adaptation分類
    • よくまとまっていて非常にわかりやすい

  • Regression layers が一番効いている
    • “This is because in the autonomous driving scenario, there exists a large difference between categories such as car and pedestrian, the source model can already achieve high classification accuracy and the main factor for limiting the detection accuracy is the inaccurate regression results. Based on the observation, in this work, we focus on studying how to only adapt the parameters of regression layers and make better regression.”

  • Whole architecture

  • pipeline
    • Noise-consistency pseudo label generation
    • Confidence-guided regression refinement

Experiment

  • model: SECOND-IoU
  • Result