Research map / Video / Video segmentation and tracking
Recurrent Graph Neural Networks for Video Instance Segmentation
Emil Brissman, Joakim Johnander, Martin Danelljan, Michael Felsberg
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 |
|---|---|---|---|
| use full graph neural network information flow over all tracks and detections (ablation) | track model · YouTubeVIS val (ablation, ResNet50) | mAP: 28.6 -> 35.3 | Improved |
| predict association with the GNN instead of a hand-made rule (ablation) | track model · YouTubeVIS val (ablation, ResNet50) | mAP: 29.2 -> 35.3 | Improved |
| add appearance model (ablation) | track model · YouTubeVIS val (ablation, ResNet50) | mAP: 34.5 -> 35.3 | Improved |
| learn appearance variance instead of constant (ablation) | track model · YouTubeVIS val (ablation, ResNet50) | mAP: 33.0 -> 35.3 | Improved |
| Gaussian appearance with learnable update instead of LSTM update (ablation) | track model · YouTubeVIS val (ablation, ResNet50) | mAP: 34.0 -> 35.3 | Improved |
| treat appearance separately instead of baking into embedding (ablation) | track model · YouTubeVIS val (ablation, ResNet50) | mAP: 34.4 -> 35.3 | Improved |
| score tracks from track embeddings instead of averaging detection confidence (ablation) | track model · YouTubeVIS val (ablation, ResNet50) | mAP: 31.5 -> 35.3 | Improved |
| score tracks from embeddings instead of averaging YOLACT scores (ablation) | track model · YouTubeVIS val (ablation, ResNet50) | mAP: 30.7 -> 35.3 | Improved |
| add LSTM-like gating to the recurrent connection (ablation) | track model · YouTubeVIS val (ablation, ResNet50) | mAP: Diverges -> 35.3 | Improved |
| use full LSTM-like gating instead of a single gate with tanh (ablation) | track model · YouTubeVIS val (ablation, ResNet50) | mAP: 31.5 -> 35.3 | Improved |
| replace ResNet50 with ResNet101 backbone | track model · YouTubeVIS val (ablation, ResNet50) | mAP: 35.3 -> 37.7 | Improved |
| use sigmoid-gated node updates instead of 2-layer MLP (ablation) | GNN · YouTubeVIS val (ablation, ResNet50) | mAP: 34.1 -> 35.3 | 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.
- 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
- End-to-End Video Instance Segmentation via Spatial-Temporal Graph Neural Networks (ICCV 2021 · similarity 0.93)
- Learning Video Object Segmentation With Visual Memory (ICCV 2017 · similarity 0.90)
- Video Object Segmentation with Episodic Graph Memory Networks (ECCV 2020 · similarity 0.89)
- Recurrent Instance Segmentation (ECCV 2016 · similarity 0.89)
- MaskRNN: Instance Level Video Object Segmentation (NeurIPS 2017 · similarity 0.89)
- Minimizing Labeled, Maximizing Unlabeled: An Image-Driven Approach for Video Instance Segmentation (CVPR 2025 · similarity 0.87)
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.
