Recoverability-Aware Few-Shot Point Cloud Segmentation for Outdoor Scenes

An initial research release for few-shot semantic segmentation of outdoor point clouds. The repository focuses on feature recovery, reliability-aware supervision, and controlled adaptation under limited annotations.

Introduction

Few-shot point cloud segmentation aims to recognize novel semantic classes from only a small number of annotated support scenes. This setting is particularly relevant outdoors, where dense point-level labels are expensive to acquire and where scene geometry, point density, object scale, and class frequency can change substantially across locations and sensors.

The central difficulty is that a small support set provides an incomplete description of each target class. A model can fit the observed support points while failing to transfer to query regions with different local structures or feature statistics. Foreground scarcity and class imbalance further amplify this support-query gap, making unrestricted adaptation prone to memorization and overly conservative adaptation prone to underfitting.

This project studies a recoverability-aware prototype framework. Support features are aggregated into foreground and background prototypes for query prediction. A bounded Residual+Norm adapter then performs a small task-specific correction while preserving the pretrained point representation. In parallel, a feature-availability score estimates support-query discrepancy and support scarcity.

The feature-availability score does not replace distillation. Instead, it enters a local supervision gate together with student need, teacher reliability, and geometric recoverability:

gate = studentNeed * teacherReliability * recoverability * missingFeature

This factorization separates two questions: where feature correction is needed, and where the available supervisory signal is reliable enough to apply. If any factor is weak, the gate suppresses strong local distillation; when all factors are strong, supervision is concentrated on regions where recovery is both necessary and feasible.

Core Components

• Prototype-based support/query segmentation with foreground-oriented evaluation.

• Bounded Residual+Norm feature correction:

  f' = normalize(f + alpha * MLP(LayerNorm(f))),  0 <= alpha <= alpha_max
  

• Feature-availability scoring from support-query feature discrepancy and support scarcity.

• Recoverability-aware local distillation controlled by a multiplicative gate.

• Generalization controls including differential learning rates, mild geometry augmentation, dropout, label smoothing, gradient clipping, best-only checkpointing, and early stopping.

Repository Layout

code/       Minimal custom training, evaluation, episode-generation, and OPD modules
docs/       Introduction outline and claim-evidence audit
LICENSE     MIT License

Experimental tables, raw logs, checkpoints, datasets, and generated episodes are intentionally excluded from this initial release. They will be added after the evaluation protocol and reporting format are finalized.

Requirements

The released files are designed to be placed on top of an existing compatible point-cloud segmentation environment. Core dependencies include:

• Python 3.10+
• PyTorch with CUDA support
• NumPy
• h5py
• flash-attention support required by the host point encoder

The complete host framework, pretrained weights, datasets, and third-party CUDA dependencies are not redistributed here.

Quick Start

1. Prepare a compatible point-cloud segmentation environment.
2. Copy the contents of code/ into the corresponding repository paths.
3. Generate N-way K-shot support/query episodes with the episode-generation script.
4. Set REPO_ROOT, DATA_ROOT, PYTHON_BIN, GPU_ID, and SOURCE_CKPT for your environment.
5. Launch the reference configuration:

bash code/run_isprs_s0_3w5s_v2.sh

The reference launcher performs periodic full-episode evaluation and retains only the best checkpoint according to foreground mean IoU.

Research Status

This repository is an early research release. The implementation is provided to make the proposed mechanism inspectable and reproducible, but it should not yet be interpreted as a finalized benchmark package or a universal performance claim.

License

The custom material in this repository is released under the MIT License. Third-party dependencies and host-framework components remain subject to their own licenses; see THIRD_PARTY_NOTICES.md.