Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Invariant Inter-pixel Correlations for Superpixel Generation
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 CDS, a deep superpixel method that uses auxiliary gradient-map modality and disentangles invariant inter-pixel correlations from dataset style noise via local-grid alignment and global-style mutual information minimization.
Problem
Deep superpixel algorithms learn features that are sensitive to dataset statistics (color distribution, semantics), hurting generalization to unseen/open-world images and boundary adherence.
Approach
Construct auxiliary modalities (HSV, LAB, gradient map) homologous to RGB but with style shifts; use a shared channel-wise content selective gate, a local-grid KL alignment of content features across modalities, and a variational mutual-information upper bound to minimize style MI, so that only invariant content is used for superpixel association prediction at inference from RGB alone.
Key results
- Trained only on BSDS500, CDS outperforms SCN and AINet on ASA, BR-BP and UE across BSDS, NYUv2, KITTI and VOC2012 (paper reports consistent gains across domains).
- Runtime of about 6 ms on GPU for ~600 superpixels on NYUv2, comparable to SCN (5 ms) and faster than AINet (29 ms).
- Downstream DeepLab + bilateral inception on Pascal VOC2012 reaches 79.02 mIoU with CDS superpixels versus 78.90 for SCN and 78.96 for AINet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CDS superpixels instead of AINet superpixels | DeepLab with the BI module · Pascal VOC2012 reduced validation set, 600 superpixels | IoU: 78.96 -> 79.02 | Helped |
| Use CDS instead of SCN for superpixel generation | NYUv2 runtime comparison · NYUv2, about 600 superpixels, 608×448 | runtime (ms): 5 -> 6 | Hurt |
| Use gradient map rather than HSV or LAB as the auxiliary modality (ablation) | CDS · BSDS and NYUv2 | ASA | Helped |
| Use only feature alignment constraints, without style mutual information minimization (ablation) | CDS · BSDS and NYUv2 | ASA | Hurt |
| Use only style mutual information minimization, without feature alignment constraints (ablation) | CDS · BSDS and NYUv2 | ASA | Hurt |
| Replace RGB training inputs with randomly selected auxiliary modalities as data augmentation (ablation) | Single-modality superpixel model · BSDS and NYUv2 | ASA | Hurt |
| Replace batch normalization layers with instance normalization (ablation) | CDS model structure · BSDS and NYUv2 | ASA | Hurt |
| Apply whitening loss to features after feature extraction (ablation) | CDS model structure (IW) · BSDS and NYUv2 | ASA | Mixed |
| Apply group instance whitening to features after feature extraction (ablation) | CDS model structure (GIW) · BSDS and NYUv2 | ASA | Mixed |
Surprising findings
- Instance Normalization, a common style-removal technique, performed worse than the baseline on generalization in this setting.
- Using the auxiliary modalities as plain data augmentation caused a severe performance decline.
Limitations
- Inference cost grows when generating a larger number of superpixels, a shared limitation of neighborhood-classification superpixel methods.
- Compactness (CO) is slightly lower than SCN and SLIC.
Experimental setup
Datasets: BSDS500
Benchmarks: BSDS500 NYUv2 KITTI Pascal VOC2012
Metrics: ASA BR-BP UE CO runtime (ms) mIoU (downstream DeepLab + bilateral inception)
Baselines: SLIC ERS ETPS SEAL SSN SCN AINet IN IW GIW
Builds on: SCN SSN instance normalization group-wise instance whitening
Closest papers by meaning
- Superpixel Segmentation With Fully Convolutional Networks (CVPR 2020 · similarity 0.86)
- Breaking Alignment Barriers: TPS-Driven Semantic Correlation Learning for Alignment-Free RGB-T Salient Object Detection (AAAI 2026 · similarity 0.85)
- Differentiable Laplacian Matrix Guided Superpixel Segmentation (CVPR 2026 · similarity 0.85)
- Soft Superpixel Neighborhood Attention (NeurIPS 2024 · similarity 0.83)
- PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining (NeurIPS 2026 · similarity 0.83)
- SeD: Semantic-Aware Discriminator for Image Super-Resolution (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.