ProtoFlow: Mitigating Forgetting in Class-Incremental Remote Sensing Segmentation with Time-Aware Prototype Dynamics

Forum topic · 小凯 · 2026-04-06

Summary

ProtoFlow is a time-aware prototype dynamics framework for class-incremental and domain-incremental remote sensing segmentation, introduced by Jiekai Wu, Rong Fu, Chuangqi Li, and colleagues on arXiv (2604.03212). The authors observe that real-world remote sensing deployment is inherently continual: new semantic categories emerge over time, while acquisition conditions shift across seasons, cities, and sensors. Existing incremental approaches often treat training steps as isolated updates, leaving representation drift and catastrophic forgetting insufficiently controlled. ProtoFlow addresses this by modeling class prototypes as trajectories and learning their evolution through an explicit temporal vector field. It jointly enforces low-curvature motion and inter-class separation, stabilizing prototype geometry throughout incremental learning. Experiments on standard class- and domain-incremental remote sensing benchmarks show consistent gains over strong baselines, improving overall mIoU by 1.5-2.0 points while reducing forgetting. This forum post summarizes the paper for zhichai.net readers, including the abstract and arXiv link.

Paper Overview

Research Area: Computer Vision (CV) Authors: Jiekai Wu, Rong Fu, Chuangqi Li, et al. Published: 2026-04-03 arXiv: 2604.03212

Abstract

Remote sensing segmentation in real deployment is inherently continual: new semantic categories emerge, and acquisition conditions shift across seasons, cities, and sensors. Despite recent progress, many incremental approaches still treat training steps as isolated updates, which leaves representation drift and forgetting insufficiently controlled.

The authors present ProtoFlow, a time-aware prototype dynamics framework that models class prototypes as trajectories and learns their evolution with an explicit temporal vector field. By jointly enforcing low-curvature motion and inter-class separation, ProtoFlow stabilizes prototype geometry throughout incremental learning.

Key Results

Consistent gains over strong baselines on standard class-incremental and domain-incremental remote sensing benchmarks
Improved overall mIoU (mIoU_all) by 1.5-2.0 points
Reduced forgetting of previously learned classes

Why It Matters

Continual learning is essential for deployed remote sensing systems, where labels for new categories accumulate over time and sensor/seasonal/domain shifts degrade static models. ProtoFlow's prototype-trajectory formulation offers a principled way to control representation drift during incremental updates.

---

*Auto-collected on 2026-04-06*