Research map / Video / Video segmentation and tracking
Labelling unlabelled videos from scratch with multi-modal self-supervision
Yuki Asano, Mandela Patrick, Christian Rupprecht, Andrea Vedaldi
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 |
|---|---|---|---|
| Train clusters and representations jointly with SeLaVi instead of applying k-means to pretrained representations | VGG-Sound video clustering · VGG-Sound | NMI: 48.5 -> 55.9 | Improved |
| Train clusters and representations jointly with SeLaVi instead of applying k-means to pretrained representations | AVE video clustering · AVE | NMI: 56.3 -> 66.2 | Improved |
| Train clusters and representations jointly with SeLaVi instead of applying k-means to pretrained representations | Kinetics video clustering · Kinetics | NMI: 48.9 -> 27.1 | Hurt |
| Train clusters and representations jointly with SeLaVi instead of applying k-means to pretrained representations | Kinetics-Sound video clustering · Kinetics-Sound | NMI: 47.5 -> 47.5 | No clear effect |
| Add audio to visual input (ablation) | SeLaVi · VGG-Sound | NMI: 52.8 -> 55.9 | Improved |
| Add visual input to audio-only clustering (ablation) | SeLaVi · VGG-Sound | NMI: 47.5 -> 55.9 | Improved |
| Use modality-splicing and shared multi-modal labels instead of feature concatenation (ablation) | SeLaVi · VGG-Sound | NMI: 24.7 -> 48.8 | Improved |
| Add modality alignment (ablation) | SeLaVi · VGG-Sound | NMI: 50.9 -> 51.1 | Improved |
| Add Gaussian marginals (ablation) | SeLaVi · VGG-Sound | NMI: 51.5 -> 51.1 | Hurt |
| Add decorrelated heads (ablation) | SeLaVi · VGG-Sound | NMI: 49.9 -> 51.1 | Improved |
| Vary the number of clusters from 309 to 256 or 619 (ablation) | SeLaVi · VGG-Sound | NMI: 56.7 -> 56.8; 56.7 -> 56.9 | No clear effect |
| Increase the number of clusters from 309 to 619 (ablation) | SeLaVi · VGG-Sound | mean maximal purity: 38.0 -> 42.7 | Improved |
Closest papers by meaning
- Unsupervised Semantic Scene Labeling for Streaming Data (CVPR 2017 · similarity 0.88)
- Multiview Pseudo-Labeling for Semi-Supervised Learning From Video (ICCV 2021 · similarity 0.87)
- The Pursuit of Human Labeling: A New Perspective on Unsupervised Learning (NeurIPS 2023 · similarity 0.87)
- UniViT: Unifying Image and Video Understanding in One Vision Encoder (NeurIPS 2025 · similarity 0.87)
- Minimizing Labeled, Maximizing Unlabeled: An Image-Driven Approach for Video Instance Segmentation (CVPR 2025 · similarity 0.86)
- Weakly-Supervised Audio-Visual Video Parsing with Prototype-based Pseudo-Labeling (CVPR 2024 · similarity 0.85)
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.
