Research map / Detection and segmentation / Weakly supervised detection and segmentation
Point-supervised Panoptic Segmentation via Estimating Pseudo Labels from Learnable Distance
Jing Li, Junsong Fan, Zhaoxiang Zhang
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 |
|---|---|---|---|
| estimate pseudo labels with a learnable distance branch instead of rule-based Dijkstra distances | point-supervised panoptic segmentation · PASCAL VOC val (P1, Swin-L) | PQ: 61.0 -> 68.5 | Improved |
| same, with ten-point labels on COCO | EPLD R50 · COCO val (P10) | PQ: 35.2 -> 41.3 | Improved |
| aggregate anchor queries from fusion feature rather than learned position embedding (ablation) | EPLD · PASCAL VOC val (P1) | PQ: 51.4 -> 56.6 | Improved |
| add class embedding (E_cls) and class loss (L_cls) for query enhancing (ablation) | EPLD · PASCAL VOC val (P1) | PQ: 42.1 -> 56.6 | Improved |
| add self-training loss with last-layer pseudo labels (ablation) | EPLD · PASCAL VOC val (P1) | PQ: 15.6 -> 56.6 | Improved |
| add color-prior loss (ablation) | EPLD · PASCAL VOC val (P1) | PQ: 46.5 -> 56.6 | Improved |
| use iterative decoder layers (6) with re-aggregated queries instead of 1 or 2 (ablation) | EPLD · PASCAL VOC val (P1) | PQ: 3.5 -> 56.6 (1 layer vs 6) | Improved |
| increase decoder layers from 6 to 7 (ablation) | EPLD · PASCAL VOC val (P1) | PQ: 56.6 -> 55.9 | Hurt |
| train panoptic branch with distance-branch pseudo labels instead of point labels (ablation) | Panoptic SegFormer head · PASCAL VOC val (P1) | PQ: 19.5 -> 56.6 | Improved |
| use multi-scale features F1, F2, F3 in fusion and decoder instead of fewer scales (ablation) | EPLD · PASCAL VOC val (P1) | PQ: 55.6 -> 56.6 (decoder F1 only vs full) | Improved |
| use EPLD pseudo labels instead of SAM (ViT-H) prompted with point labels | Panoptic SegFormer Swin-L · PASCAL VOC val | PQ: 60.6 -> 72.2 | Improved |
Closest papers by meaning
- Fully Data-Driven Pseudo Label Estimation for Pointly-Supervised Panoptic Segmentation (AAAI 2024 · similarity 0.94)
- Pointly-Supervised Panoptic Segmentation (ECCV 2022 · similarity 0.94)
- Point2Mask: Point-supervised Panoptic Segmentation via Optimal Transport (ICCV 2023 · similarity 0.91)
- Scene-Centric Unsupervised Panoptic Segmentation (CVPR 2025 · similarity 0.89)
- Beyond the Label Itself: Latent Labels Enhance Semi-supervised Point Cloud Panoptic Segmentation (AAAI 2024 · similarity 0.89)
- Fully Convolutional Networks for Panoptic Segmentation with Point-based Supervision (TPAMI 2022 · similarity 0.89)
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.
