Research map / Video / Video segmentation and tracking
Sound Localization from Motion: Jointly Learning Sound Direction and Camera Rotation
Ziyang Chen, Shengyi Qian, Andrew Owens
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reflection-invariant (permutation invariant) geometric loss to handle front-back ambiguity (ablation) | SLfM · HM3D-SS LibriSpeech | camera angle MAE (deg): 26.80 -> 0.77 | Improved |
| add binaural IID loss (ablation) | SLfM · HM3D-SS LibriSpeech | camera angle MAE (deg): 29.20 -> 6.99 | Improved |
| add symmetry loss on top of geometric and binaural losses (ablation) | SLfM · HM3D-SS LibriSpeech | camera angle MAE (deg): 6.99 -> 0.77 | Improved |
| pretrain features with cross-view binauralization instead of random features (frozen) (ablation) | SLfM · HM3D-SS | camera angle MAE (deg): 29.26 -> 0.77 | Improved |
| finetune features during SLfM instead of freezing (ablation) | SLfM · HM3D-SS | audio angle MAE (deg): 3.17 -> 2.77 | Improved |
| use both visual and audio features in binauralization pretext (ablation) | cross-view binauralization · HM3D-SS LibriSpeech pretext | STFT distance: 0.206 -> 0.130 | Improved |
| train pretext with 3 target views instead of 2 (ablation) | cross-view binauralization · HM3D-SS LibriSpeech linear probe | camera rotation accuracy (%): 80.0 -> 81.3 | Improved |
| predict right channel from left (L2R) instead of mono-to-binaural (ablation) | cross-view binauralization · HM3D-SS LibriSpeech linear probe | audio localization accuracy (%): 75.4 -> 78.5 audio; 81.3 -> 76.1 rotation | Mixed |
| use cross-attention U-Net instead of 2.5D U-Net for audio prediction (ablation) | cross-view binauralization · HM3D-SS LibriSpeech | downstream camera rotation accuracy (%): 80.0 -> 75.3 | Hurt |
| increase reverberation (RT60) in rendered audio (ablation) | SLfM · HM3D-SS | SLfM accuracy | Hurt |
| geometric loss weight lambda in range 3 to 8 (ablation) | SLfM · HM3D-SS | SLfM performance | No clear effect |
| use SLfM-trained visual model instead of SIFT/Superglue feature matching | SLfM · Stanford2D3D | mean rotation error (deg): 5.07 -> 1.14 (vs Superglue) | Improved |
Closest papers by meaning
- Supervising Sound Localization by In-the-wild Egomotion (CVPR 2025 · similarity 0.92)
- Binaural Audio-Visual Localization (AAAI 2021 · similarity 0.91)
- Localize to Binauralize: Audio Spatialization From Visual Sound Source Localization (ICCV 2021 · similarity 0.91)
- Binaural SoundNet: Predicting Semantics, Depth and Motion With Binaural Sounds (TPAMI 2022 · similarity 0.91)
- Cyclic Learning for Binaural Audio Generation and Localization (CVPR 2024 · similarity 0.89)
- Seeing Sound, Hearing Sight: Uncovering Modality Bias and Conflict of AI models in Sound Localization (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
