Clink! Chop! Thud! -- Learning Object Sounds from Real-World Interactions

  • 2025-10-02 17:59:52
  • Mengyu Yang, Yiming Chen, Haozheng Pei, Siddhant Agarwal, Arun Balajee Vasudevan, James Hays
  • 0

Abstract

Can a model distinguish between the sound of a spoon hitting a hardwood floorversus a carpeted one? Everyday object interactions produce sounds unique tothe objects involved. We introduce the sounding object detection task toevaluate a model's ability to link these sounds to the objects directlyinvolved. Inspired by human perception, our multimodal object-aware frameworklearns from in-the-wild egocentric videos. To encourage an object-centricapproach, we first develop an automatic pipeline to compute segmentation masksof the objects involved to guide the model's focus during training towards themost informative regions of the interaction. A slot attention visual encoder isused to further enforce an object prior. We demonstrate state of the artperformance on our new task along with existing multimodal action understandingtasks.

 

Quick Read (beta)

loading the full paper ...