Research map / Detection and segmentation / CNN features
PolarNeXt: Rethink Instance Segmentation with Polar Representation
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 PolarNeXt, a lightweight polar-representation instance segmentation framework that adds adaptive polygonal sample decisions and a union-aligned rasterization module (RMask IoU) to account for representation error.
Problem
Polar-representation instance segmentation methods are efficient but lag far behind mainstream mask- and contour-based methods in accuracy, because representation error of polygons is ignored in sample assignment and polygon assessment.
Approach
Treats polygon representation error as a missing error term: an Adaptive Polygonal Sample Decision selects positives by matching costs (not center prior) over multi-layer center sampling, and a Union-aligned Rasterization Module computes a differentiable mask IoU (RMask IoU) between predicted polygon and contour for weighting and loss.
Key results
- PolarNeXt reaches 36.1 AP on COCO test-dev with R50-FPN (36 epochs), a 4.8 AP gain over PolarMask (31.3), at 49 FPS and 32.36M params.
- On Cityscapes val, PolarNeXt gets 30.7 AP vs PolarMask 27.4 AP, with memory 566 MB vs Mask R-CNN 1968 MB.
- Ablation: URM and APSD together add 4.8 AP on COCO val (29.1 to 33.9) at 1x schedule.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace PolarMask's center-prior sample decisions with APSD and URM | PolarNeXt · COCO test-dev | mask AP: 31.3 -> 36.1; 4.8% mask AP improvement | Helped |
| replace Centerness sample weights with RMask IoU (ablation) | PolarNeXt · COCO val | AP: 29.1 -> 30.3 (+1.2) | Helped |
| add RMask IoU Loss to the loss functions alongside RMask IoU sample weighting (ablation) | PolarNeXt · COCO val | AP: 30.3 -> 31.0 (+1.9 over baseline) | Helped |
| apply APSD without URM (ablation) | PolarNeXt · COCO val | AP: 29.1 -> 28.6 (−0.5) | Hurt |
| include RMask IoU Loss in APSD matching costs (ablation) | PolarNeXt · COCO val | AP: 30.7 -> 33.9 | Helped |
| remove Polar IoU Loss (ablation) | PolarNeXt · COCO val | AP: 33.9 -> 32.8 | Hurt |
| remove Polar IoU Loss (ablation) | PolarNeXt · COCO val, small objects | APS: 16.4 -> 14.3 | Hurt |
| increase single-layer Center Sampling radius from 1.5 to 2.0 (ablation) | PolarNeXt candidate selection · COCO val | AP: 31.0 -> 30.6 (−0.4) | Hurt |
| select candidates using Center Sampling across all FPN layers (ablation) | PolarNeXt candidate selection · COCO val | AP: 31.0 -> 33.9 (+2.9) | Helped |
| use points within boxes for candidate selection across FPN layers (ablation) | PolarNeXt candidate selection · COCO val | AP: 33.1 (+2.1) | Helped |
| use points within contours for candidate selection across FPN layers (ablation) | PolarNeXt candidate selection · COCO val | AP: 33.0 (+2.0) | Helped |
| replace PolarMask with PolarNeXt | R50-FPN instance segmentation · Cityscapes val | mask AP: 27.4 -> 30.7 | Helped |
Surprising findings
- Applying APSD without the RMask IoU loss in matching costs fails or even hurts performance (28.6 AP vs 29.1 baseline).
- Directly expanding center sampling radius on a single FPN layer decreases AP by 0.4.
Limitations
- Accuracy remains below the strongest two-stage methods such as HTC (38.7 AP) and Cascade R-CNN (38.4 AP) on COCO test-dev.
- Relies on RMask IoU and mask rasterization at a fixed 64x64 resolution, with a sharpness hyperparameter tau.
Experimental setup
Datasets: MS COCO Cityscapes ImageNet
Benchmarks: MS COCO test-dev MS COCO val Cityscapes val
Metrics: mask AP AP50 AP75 APS APM APL boundary AP FPS parameters FLOPs
Baselines: PolarMask ESE-Seg Mask R-CNN Cascade R-CNN HTC CondInst SOLOv2 SparseInst Deep Snake E2EC PolySnake DANCE
Builds on: PolarMask FCOS Polar IoU Loss Center Sampling
Closest papers by meaning
- PolarMask: Single Shot Instance Segmentation With Polar Representation (CVPR 2020 · similarity 0.90)
- Instance Segmentation With Mask-Supervised Polygonal Boundary Transformers (CVPR 2022 · similarity 0.87)
- Insightful Instance Features for 3D Instance Segmentation (CVPR 2025 · similarity 0.84)
- EFDTR: Learnable Elliptical Fourier Descriptor Transformer for Instance Segmentation (ICML 2025 · similarity 0.83)
- RGB-to-Polarization Estimation: A New Task and Benchmark Study (NeurIPS 2025 · similarity 0.83)
- Spherical Mask: Coarse-to-Fine 3D Point Cloud Instance Segmentation with Spherical Representation (CVPR 2024 · similarity 0.83)
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.