Research map / Video / Video segmentation and tracking
Learning What to Learn for Video Object Segmentation
Goutam Bhat, Felix Järemo Lawin, Martin Danelljan, Andreas Robinson, Michael Felsberg, Luc Van Gool, Radu Timofte
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 end-to-end training through the few-shot learner (ablation) | Few-shot VOS · YouTube-VOS 2019 train split, 300-video validation | J: 74.5 -> 77.5 | Improved |
| add learned label generator (ablation) | Few-shot VOS · YouTube-VOS 2019 train split, 300-video validation | J: 77.5 -> 78.9 | Improved |
| add learned importance weight predictor (ablation) | Few-shot VOS · YouTube-VOS 2019 train split, 300-video validation | J: 78.9 -> 79.8 | Improved |
| initialize backbone with Mask-RCNN weights instead of ImageNet (ablation) | Few-shot VOS · YouTube-VOS 2019 train split, 300-video validation | J: 78.6 -> 79.8 | Improved |
| use Lovasz loss instead of binary cross-entropy (ablation) | Few-shot VOS · YouTube-VOS 2019 train split, 300-video validation | J: 79.2 -> 79.8 | Improved |
| use long training instead of short training (ablation) | Few-shot VOS · YouTube-VOS 2019 train split, 300-video validation | J: 79.8 -> 81.2 | Improved |
| use mask encoding with D>=4 channels instead of 1 (ablation) | Few-shot VOS · YouTube-VOS 2019 train split, 300-video validation | J: +2.0 | Improved |
| perform one steepest-descent update per frame instead of none (ablation) | Few-shot VOS · YouTube-VOS 2019 train split, 300-video validation | J: more than 5 points | Improved |
| use more than 3 steepest-descent iterations per frame (ablation) | Few-shot VOS · YouTube-VOS 2019 train split, 300-video validation | J: 81.2 -> slight degradation | Hurt |
| evaluate on local search region instead of full frame (ablation) | Few-shot VOS · YouTube-VOS 2019 train split, 300-video validation | J: 80.2 -> 81.2 | Improved |
| use update rate 0.1 for global importance weights (ablation) | Few-shot VOS · YouTube-VOS 2019 train split, 300-video validation | J: 81.2 | Improved |
| increase maximum training set size K_max up to 16 (ablation) | Few-shot VOS · YouTube-VOS 2019 train split, 300-video validation | J | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- AGSS-VOS — AGSS-VOS: Attention Guided Single-Shot Video Object Segmentation (ICCV 2019) · better on 3 of 3 shared results
- FEELVOS — FEELVOS: Fast End-To-End Embedding Learning for Video Object Segmentation (CVPR 2019) · better on 1 of 1 shared result
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
- Learning Fast and Robust Target Models for Video Object Segmentation (CVPR 2020 · similarity 0.93)
- Make One-Shot Video Object Segmentation Efficient Again (NeurIPS 2020 · similarity 0.93)
- Low-Shot Video Object Segmentation (TPAMI 2025 · similarity 0.93)
- Two-Shot Video Object Segmentation (CVPR 2023 · similarity 0.93)
- Point-VOS: Pointing Up Video Object Segmentation (CVPR 2024 · similarity 0.91)
- Temporal Transductive Inference for Few-Shot Video Object Segmentation (IJCV 2025 · similarity 0.91)
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.
