Research map / 3D vision / Gaussian splatting
Chorus: Multi-Teacher Pretraining for Holistic 3D Gaussian Scene Encoding
Yue Li, Qi Ma, Runyi Yang, Mengjiao Ma, Bin Ren, Nikola Popovic, Nicu Sebe, Theo Gevers, Luc Van Gool, Danda Pani Paudel, Martin R. Oswald
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 DINO teacher to the language teacher (ablation) | Chorus (3DGS) · ScanNet++ v2-trained, zero-shot seg on ScanNet++ val | f-mIoU: 27.1 -> 29.4 | Improved |
| add PE-Spatial teacher on top of Lang+DINO (ablation) | Chorus (3DGS) · InteriorGS zero-shot seg | f-mIoU: 9.3 -> 11.4 | Improved |
| add PE-Spatial teacher on top of Lang+DINO (ablation) | Chorus (3DGS) · ScanNet++ val zero-shot seg | f-mIoU: 29.4 -> 29.6 | No clear effect |
| introduce PE-Spatial only in second half of training instead of from the start (ablation) | Chorus pretraining · ScanNet++ val zero-shot seg | segmentation score | Improved |
| use SmoothL1 loss, 3DGS-aware augmentations and instance contrastive term (ablation) | Chorus pretraining · ScanNet++ val zero-shot seg | segmentation score | Improved |
| replace SceneSplat with Chorus 3D backbone in GaussianVLM | GaussianVLM · Nr3D | Sim: 20.8 -> 22.5 | Improved |
| replace SceneSplat with Chorus in GaussianVLM | GaussianVLM · ScanQA | METEOR: 22.9 -> 22.5 (M); 14.4 -> 14.8 (EM1) | Mixed |
| Chorus multi-teacher pretraining vs Sonata self-supervised, linear probing | point-cloud encoder · ScanNet200 val | mIoU: 28.8 -> 36.0 | Improved |
| Chorus vs Sonata in linear-probe instance segmentation on ScanNet | point-cloud encoder · ScanNet val instance seg | mAP25: 72.6 -> 66.6 | Hurt |
| Chorus vs Sonata with decoder probing on ScanNet semantic seg | point-cloud encoder · ScanNet val | mIoU: 77.3 -> 75.0 | Hurt |
| add render-and-distill adaptation on 100 extra InteriorGS scenes | Chorus · InteriorGS linear probing | mIoU: +2.7% mIoU | Improved |
| Chorus pretraining as input-noise-robust features | Chorus point-cloud variant · ScanNet++ instance retrieval PC to noisy PC | R@1: 79.8% -> 85.4% | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SceneSplat — SceneSplat++: A Large Dataset and Comprehensive Benchmark for Language Gaussian Splatting (NeurIPS 2025) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SceneSplat++: A Large Dataset and Comprehensive Benchmark for Language Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- Dr. Splat: Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration (CVPR 2025 · similarity 0.88)
- UniGS: Unified Language-Image-3D Pretraining with Gaussian Splatting (ICLR 2025 · similarity 0.88)
- Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields (CVPR 2024 · similarity 0.88)
- UniPre3D: Unified Pre-training of 3D Point Cloud Models with Cross-Modal Gaussian Splatting (CVPR 2025 · similarity 0.88)
- FHGS: Feature-Homogenized Gaussian Splatting (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.
