Research map / Detection and segmentation / Open-set and few-shot detection
Unsupervised Discovery of the Long-Tail in Instance Segmentation Using Hierarchical Self-Supervision
Zhenzhen Weng, Mehmet Giray Ogut, Shai Limonchik, Serena Yeung
Check 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 hierarchical self-supervised mask embeddings without long-tail box annotations | Instance segmentation trained with COCO masks · LVIS (COCO to LVIS) | mAP: 0.056 -> 0.109 | Improved |
| Use hierarchical self-supervised mask embeddings without long-tail box annotations | Instance segmentation trained with VOC masks · COCO non-VOC categories | mAP: 0.238 -> 0.327 | Improved |
| Initialize the mask proposal backbone with ImageNet pretraining rather than random initialization (ablation) | Ours · LVIS (COCO to LVIS) | mAP: 0.096 -> 0.109 | Improved |
| Increase proposals per image from 20 to 50 (ablation) | Ours · LVIS | mAP: 0.0318 -> 0.1086 | Improved |
| Increase proposals per image from 50 to 100 (ablation) | Ours · LVIS | mAP: 0.1086 -> 0.1083 | No clear effect |
| Set triplet-loss margin α to 0.2 rather than 0.1 (ablation) | Ours · LVIS | mAP: 0.0806 -> 0.1086 | Improved |
| Increase triplet-loss margin α from 0.2 to 0.5 (ablation) | Ours · LVIS | mAP: 0.1086 -> 0.0987 | Hurt |
| Set mask-loss weight β to 0.1 rather than 0.2 (ablation) | Ours · LVIS | mAP: 0.1086 -> 0.0526 | Hurt |
| Set mask-loss weight β to 0.5 rather than 0.2 (ablation) | Ours · LVIS | mAP: 0.1086 -> 0.0614 | Hurt |
| Add the mask loss to distinguish foreground from background (ablation) | Ours · LVIS | mAP: 0.0689 -> 0.1086 | Improved |
| Add the object loss to bring overlapping proposals together and separate non-overlapping proposals (ablation) | Ours · LVIS | mAP: 0.0374 -> 0.1086 | Improved |
| Add the hierarchical loss to encode parent-child relations (ablation) | Ours · LVIS | mAP: 0.0846 -> 0.1086 | 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.
- ShapeMask — ShapeMask: Learning to Segment Novel Objects by Refining Shape Priors (ICCV 2019) · better on 2 of 2 shared results
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
- The Devil is in Classification: A Simple Framework for Long-tail Instance Segmentation (ECCV 2020 · similarity 0.92)
- Unsupervised Semantic Segmentation with Self-supervised Object-centric Representations (ICLR 2023 · similarity 0.90)
- Learning with Free Object Segments for Long-Tailed Instance Segmentation (ECCV 2022 · similarity 0.89)
- SOLO: Segmenting Objects by Locations (ECCV 2020 · similarity 0.89)
- Object Discovery and Representation Networks (ECCV 2022 · similarity 0.89)
- Training-Free Dataset Pruning for Instance Segmentation (ICLR 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.
