Research map / Detection and segmentation / CNN features
Training-Free Dataset Pruning for Instance Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TFDP, a training-free dataset pruning method for instance segmentation that ranks images by mask boundary complexity (perimeter-to-area ratio) with scale and class normalization, avoiding any model training.
Problem
Dataset pruning for instance segmentation is underexplored; existing pruning methods are designed for classification, require costly model training to score samples, and do not handle pixel-level labels, varying instance areas, or class imbalance.
Approach
Score each image by a model-independent Shape Complexity Score (perimeter-to-area ratio of instance masks), made scale-invariant by normalizing against a circle and class-balanced by normalizing instance scores within each class across images, then sum per image and keep top-ranked images.
Key results
- On COCO with Mask R-CNN at 50% pruning, TFDP reaches 53.4% mask AP50 vs. 51.2% for adapted EL2N-style baselines' best (+2.4 over CCS/Entropy-level), outperforming random by 2.4-3.1 points at 40-50% pruning.
- Selection time on COCO is 0.014 h versus about 12-21 h for adapted baselines, an average speedup of 1349x.
- On VOC, at 40% and 50% pruning TFDP reaches 36.2% and 33.4% mask AP50, exceeding random pruning by 7.2 and 9.7 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TFDP instead of random pruning | Mask R-CNN · COCO, 50% pruning | mask AP50: 52.3 -> 53.4 | Helped |
| Use training-free TFDP ranking instead of model-based EL2N ranking | Mask R-CNN · COCO | ranking time: 12.37 h -> 0.014 h | Helped |
| Use TFDP-selected data instead of random selection | Mask R-CNN · VOC 2012, 50% pruning | mask AP: 23.7 -> 33.4 | Helped |
| Use TFDP-selected data instead of random selection | Mask R-CNN · Cityscapes, 50% pruning | mask AP: 16.9 -> 19.4 | Helped |
| Use TFDP-selected data instead of CCS-selected data | SOLO-v2 · COCO, 50% pruning | AP50: 52.1 -> 52.3 | Helped |
| Use TFDP-selected data instead of AUM-selected data | QueryInst · COCO, 50% pruning | AP50: 54.0 -> 55.0 | Helped |
| Add SI-SCS to SCS (ablation) | TFDP · VOC 2012, 50% pruning | mask AP: 28.8 -> 32.4 | Helped |
| Add CB-SCS to SCS (ablation) | TFDP · VOC 2012, 50% pruning | mask AP: 28.8 -> 30.7 | Helped |
| Use both SI-SCS and CB-SCS instead of SCS alone (ablation) | TFDP · VOC 2012, 50% pruning | mask AP: 28.8 -> 33.4 | Helped |
| Add SI-SCS to SCS (ablation) | TFDP · Cityscapes, 50% pruning | mask AP: 19.1 -> 17.9 | Hurt |
| Add CB-SCS to SCS (ablation) | TFDP · Cityscapes, 50% pruning | mask AP: 19.1 -> 17.1 | Hurt |
| Use both SI-SCS and CB-SCS instead of SCS alone (ablation) | TFDP · COCO, 50% pruning | mask AP: 52.5 -> 53.4 | Helped |
Surprising findings
- Adapted classification pruning baselines often perform worse than random pruning when transferred across architectures (e.g., QueryInst at some pruning rates).
Limitations
- Evaluation limited to three datasets and a small set of architectures (Mask R-CNN, SOLO-v2, QueryInst).
- Only image-level pruning is considered; instance-level annotation pruning is out of scope.
- Shape-based score may not capture semantic difficulty of samples.
Experimental setup
Datasets: Pascal VOC 2012 Cityscapes MS COCO
Benchmarks: MS COCO Pascal VOC 2012 Cityscapes
Metrics: mAP AP50 AP75 APS APM APL bbox AP
Baselines: Random Forgetting Entropy EL2N AUM CCS
Builds on: Mask R-CNN EL2N CCS Entropy AUM
Closest papers by meaning
- BoxInst: High-Performance Instance Segmentation With Box Annotations (CVPR 2021 · similarity 0.89)
- A Generative Approach at the Instance-Level for Image Segmentation Under Limited Training Data Conditions (Student Abstract) (AAAI 2025 · similarity 0.88)
- The Devil is in the Object Boundary: Towards Annotation-free Instance Segmentation using Foundation Models (ICLR 2024 · similarity 0.88)
- P2Seg: Pointly-supervised Segmentation via Mutual Distillation (ICLR 2024 · similarity 0.87)
- Exploring Learning Complexity for Efficient Downstream Dataset Pruning (ICLR 2025 · similarity 0.87)
- WISH: Weakly Supervised Instance Segmentation using Heterogeneous Labels (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.