Research map / Video / Video segmentation and tracking
TarViS: A Unified Approach for Target-Based Video Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| jointly train across tasks and datasets instead of task-specific training (ablation) | TarViS · VIS, VPS, VOS, and PET benchmarks | benchmark performance: better on 4/7 benchmarks, worse on 2/7, and matches performance on 1/7 | Mixed |
| add semantic queries for class representation (ablation) | TarViS for VIS · YouTube-VIS 2021 and OVIS | mAP: 44.7 vs. 46.3; 29.8 vs. 31.5 | Helped |
| add the Temporal Neck's inter-frame feature interactions (ablation) | TarViS · YouTube-VIS 2021, OVIS, KITTI-STEP, CityscapesVPS, VIPSeg, DAVIS, and BURST | benchmark-specific segmentation score: 42.8 vs. 48.3; 22.3 vs. 31.1; 0.69 vs. 0.70; 51.2 vs. 53.3; 28.9 vs. 33.5; 78.7 vs. 82.0; 30.3 vs. 30.9 | Helped |
| add background queries to model non-target pixels (ablation) | TarViS for VOS and PET · DAVIS and BURST | J&F / HOTA_all: 78.0 -> 81.5; 25.9 -> 29.2 | Helped |
| use four object queries per VOS target instead of one (ablation) | TarViS for VOS and PET · DAVIS and BURST | J&F / HOTA_all: 80.6 -> 81.5; 28.2 -> 29.2 | Helped |
| train jointly across tasks instead of VIS-only training (ablation) | TarViS for VIS · YouTube-VIS 2021 | mAP: 46.3 vs. 48.3 | Helped |
| train jointly across tasks instead of VIS-only training (ablation) | TarViS for VIS · OVIS | mAP: 31.5 vs. 31.1 | Hurt |
| train jointly across tasks instead of VPS-only training (ablation) | TarViS for VPS · KITTI-STEP | STQ: 0.70 vs. 0.70 | No clear effect |
| train jointly across tasks instead of VPS-only training (ablation) | TarViS for VPS · CityscapesVPS and VIPSeg | VPQ: 49.7 vs. 53.3; 32.4 vs. 33.5 | Helped |
| train jointly across tasks instead of VOS + PET-only training (ablation) | TarViS for VOS and PET · DAVIS | J&F: 81.1 vs. 82.0 | Helped |
| train jointly across tasks instead of VOS + PET-only training (ablation) | TarViS for VOS and PET · BURST | HOTA_all: 34.7 vs. 30.9 | Hurt |
Closest papers by meaning
- MinVIS: A Minimal Video Instance Segmentation Framework without Video-based Training (NeurIPS 2022 · similarity 0.88)
- SyncVIS: Synchronized Video Instance Segmentation (NeurIPS 2024 · similarity 0.88)
- STEP: Segmenting and Tracking Every Pixel (NeurIPS 2021 · similarity 0.87)
- Learning Spatial-Semantic Features for Robust Video Object Segmentation (ICLR 2025 · similarity 0.86)
- Minimizing Labeled, Maximizing Unlabeled: An Image-Driven Approach for Video Instance Segmentation (CVPR 2025 · similarity 0.86)
- Learning Segmentation from Point Trajectories (NeurIPS 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.