Research map / Detection and segmentation / Open-set and few-shot detection

In2NeCT: Inter-class and Intra-class Neural Collapse Tuning for Semantic Segmentation of Imbalanced Remote Sensing Images

AAAI 2025 · Detection and segmentation › Open-set and few-shot detection · approach: Remote sensing segmentation and change · 2 citations (OpenAlex)

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 In2NeCT, which adds inter-class (fixed simplex-ETF classifier with cross-entropy) and balanced intra-class (cosine-similarity) neural-collapse regularization branches to standard RSI segmentation models to improve minor-class performance on imbalanced remote sensing datasets.

Problem

Semantic segmentation of remote sensing images suffers from multi-scale and inconsistently distributed objects, causing data imbalance that degrades performance on minor classes; the paper asks what disrupts last-layer representation structure and how to obtain better representations.

Approach

Extends the neural collapse phenomenon to RSI segmentation: uses a frozen simplex ETF classifier and a cross-entropy inter-class loss to separate class means, plus a class-averaged cosine intra-class loss to cohere pixel features, added as training-only branches that are discarded at inference.

Key results

• On LoveDA, In2NeCT raises UNetFormer mIoU from 52.4 to 54.3 (+1.9) and reports +2.1 mIoU on MANet and +0.8 on LSKNet.
• On ISPRS Vaihingen and Potsdam, In2NeCT improves mIoU by 0.83-1.83 and 0.91-1.11 points across the three backbones.
• Training time increases about 10.5% (1.53s to 1.69s per image) with no added inference cost, since the auxiliary branches are discarded at test time.

Measured effects reported in the paper

Surprising findings

• Ablation shows the fixed ETF classifier with inter-class loss alone gives +0.5 mIoU on LoveDA while the intra-class loss alone gives +1.2, the reverse of the ordering one might expect from the emphasis on inter-class structure.

Limitations

• Evaluated only on three RSI datasets with backbones from a single family of segmentation models.
• Theoretical analysis is gradient-level and does not prove that the collapse improves generalization.
• Gains on the larger LoveDA set are modest (about 0.8-2.1 mIoU).

Experimental setup

Datasets: LoveDA ISPRS Vaihingen ISPRS Potsdam

Benchmarks: LoveDA ISPRS Vaihingen ISPRS Potsdam

Metrics: mIoU AF OA per-class IoU

Baselines: MANet Segmenter LANet DeepLabv3+ FarSeg Semantic FPN PSPNet FLANet OCRNet SwimUperNet DANet ConvNeXt

Builds on: Neural Collapse UNetFormer MANet LSKNet SupCon-style contrastive learning

Closest papers by meaning

• Sparse and Complete Latent Organization for Geospatial Semantic Segmentation (CVPR 2022 · similarity 0.87)
• Towards Open-Vocabulary Remote Sensing Image Semantic Segmentation (AAAI 2025 · similarity 0.86)
• InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.85)
• SatSynth: Augmenting Image-Mask Pairs through Diffusion Models for Aerial Semantic Segmentation (CVPR 2024 · similarity 0.85)
• SemStereo: Semantic-Constrained Stereo Matching Network for Remote Sensing (AAAI 2025 · similarity 0.84)
• Remote Sensing Vision-Language Foundation Models without Annotations via Ground Remote Alignment (ICLR 2024 · similarity 0.83)