Research map / Detection and segmentation / Open-set and few-shot detection
NTClick: Achieving Precise Interactive Segmentation With Noise-tolerant Clicks
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 NTClick, a click-based interactive segmentation method with a new noise-tolerant click type and a two-stage ECP-then-HRR workflow that refines fine structures at high resolution.
Problem
Click-based interactive segmentation becomes inefficient for thin, intricate objects because users must precisely localize narrow regions, and ViT-based methods are limited to ~1024 input resolution, losing fine detail.
Approach
Introduces a noise-tolerant click that only needs to fall near fine structures, and a two-stage pipeline: an Explicit Coarse Perception network predicts a foreground/background/uncertain (FBU) map at low resolution, then a High Resolution Refinement network uses grid and neighborhood attention to classify uncertain pixels at high resolution (up to 4096).
Key results
- On DIS5K-TE4, NTClick improves NoC@90 by 5.46 and 5-mIoU by 8.04 over the baselines (NoC@90 9.15, 5-mIoU 87.98 vs SegNext's 14.61 / 79.94 in Table 1).
- On DIS5K-VD, NTClick reaches NoC@90 7.23 and 5-mIoU 89.23, versus 10.35 / 85.69 for SegNext.
- Against Slim Scissors on DIS5K-VD, NTClick achieves 89.23 vs 82.57 5-mIoU while using click rather than click&scribble input.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use noise-tolerant clicks with the two-stage ECP→HRR workflow | NTClick · DIS5K-TE4 | 5-mIoU: 79.94 -> 87.98 | Helped |
| Use noise-tolerant clicks with the two-stage ECP→HRR workflow | NTClick · DIS5K-TE4 | NoC@90: 14.61 -> 9.15 | Helped |
| Increase ECP input resolution from 448² to 1024² with ViT-Base (ablation) | ECP · DIS5K-VD | NoC@90: 8.58 -> 7.94 | Helped |
| Increase ECP input resolution from 448² to 1024² with ViT-Base (ablation) | ECP · DIS5K-VD | 5-mIoU: 87.32 -> 88.45 | Helped |
| Scale the ECP backbone from ViT-Base to ViT-Huge at 448² (ablation) | ECP · DIS5K-VD | NoC@90: 8.58 -> 7.23 | Helped |
| Scale the ECP backbone from ViT-Base to ViT-Huge at 448² (ablation) | ECP · DIS5K-VD | 5-mIoU: 87.32 -> 89.23 | Helped |
| Remove long-range attention (ablation) | HRR · DIS5K-VD | 5-mIoU: 86.98 -> 72.17 | Hurt |
| Replace global attention with grid attention (ablation) | HRR · DIS5K-VD | 5-mIoU: 86.98 -> 88.90 | Helped |
| Replace window attention with neighborhood attention while retaining grid attention (ablation) | HRR · DIS5K-VD | 5-mIoU: 88.90 -> 89.23 | Helped |
Limitations
- Noticeable gap to ground truth remains; additional clicks yield diminishing returns after about 10 clicks.
- Resolution-scaling gains come at increased latency (0.81 s at 4096).
Experimental setup
Datasets: HQSeg-44k DIS5K ThinObject5K FSS-1000 ECSSD MSRA-10K DUT-OMRON
Benchmarks: DIS5K-VD DIS5K-TE1 DIS5K-TE2 DIS5K-TE3 DIS5K-TE4 HRSOD ThinObject-5K
Metrics: NoC@90 5-mIoU latency
Baselines: RITM CDNet FocalClick SAM SimpleClick FocSAM SAM-HQ SegNext Slim Scissors
Builds on: SimpleClick SegNext ViTMatte NAT MobileViT
Closest papers by meaning
- RClicks: Realistic Click Simulation for Benchmarking Interactive Segmentation (NeurIPS 2024 · similarity 0.89)
- TETRIS: Towards Exploring the Robustness of Interactive Segmentation (AAAI 2024 · similarity 0.88)
- Order-aware Interactive Segmentation (ICLR 2025 · similarity 0.88)
- Repurposing Stable Diffusion Attention for Training-Free Unsupervised Interactive Segmentation (CVPR 2025 · similarity 0.87)
- AGILE3D: Attention Guided Interactive Multi-object 3D Segmentation (ICLR 2024 · similarity 0.87)
- CFR-ICL: Cascade-Forward Refinement with Iterative Click Loss for Interactive Image Segmentation (AAAI 2024 · similarity 0.86)
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.