This paper proposes an active learning framework for single-photon image classification that uses imaging condition-aware synthetic augmentation and a diversity-guided uncertainty-inconsistency sampling strategy to achieve high accuracy (97% on synthetic, 90.63% on real-world data) with significantly fewer labeled samples (1.5% and 8%, respectively) compared to baselines.
Active Learning, Classification, Synthetic Data, Uncertainty Estimation, Imaging Conditions
Zili Zhang, Ziting Wen, Yiheng Qiang, Hongzhou Dong, Wenle Dong, Xinyang Li, Xiaofan Wang, Xiaoqiang Ren
Shanghai University, University of Sydney, Shanghai Institute of Technology
Generated by grok-3
Background Problem
Single-photon LiDAR enables high-precision 3D imaging in extreme environments using quantum-level photon detection, but semantic interpretation (e.g., classification) of such images is underexplored due to high annotation costs and challenges like low resolution and stochastic noise. Existing active learning methods for RGB images fail to account for variability in imaging conditions inherent to single-photon data, leading to suboptimal sample selection for labeling. This paper aims to address the gap by developing a label-efficient active learning framework for single-photon image classification, reducing annotation effort while achieving high accuracy.
Method
The proposed method introduces an active learning framework for single-photon image classification with an imaging condition-aware sampling strategy. Its core idea is to integrate synthetic augmentation into the active learning loop to model variability across imaging conditions (e.g., photon count, signal-to-background ratio) and prioritize samples where the model is uncertain and sensitive to these variations. The implementation involves: (1) generating synthetic variants of observed single-photon images under diverse conditions using a Synthetic Imaging Variants (SIV) module based on Poisson-distributed photon statistics; (2) training a ResNet-18 model on an extended dataset of observed and synthetic images; (3) selecting samples for labeling using a Diversity-guided Uncertainty-Inconsistency Sampling (DUIS) strategy, which combines margin-based uncertainty, prediction inconsistency across synthetic variants (via KL divergence), and clustering for diversity; (4) iteratively updating the labeled dataset with oracle annotations until the labeling budget is exhausted. This approach ensures efficient use of limited labels by focusing on the most informative and representative samples.
Experiment
The experiments were conducted on both synthetic and real-world single-photon datasets. The synthetic dataset, derived from RGB-D images and downsampled to mimic low-resolution single-photon imaging, included 11 categories with a 70-30 train-test split. The real-world dataset, collected using a custom single-photon LiDAR system, also spanned 11 categories with 1,100 images at 64x64 resolution. Setup involved training a ResNet-18 model with SGD over multiple active learning rounds (7 for synthetic, 10 for real-world), selecting small batches of samples per round (50 and 20, respectively). Results showed significant label efficiency: on synthetic data, the method achieved 97% accuracy with only 1.5% labeled samples, outperforming baselines like BADGE (2.5% labeled for similar accuracy) and Entropy (lowest accuracy and highest variance). On real-world data, it reached 90.63% accuracy with 8% labeled samples, a 4.51% improvement over the best baseline. The experimental design was comprehensive, comparing against uncertainty-based (Entropy, Margin), diversity-based (Coreset), and hybrid (BADGE) methods, with metrics like accuracy, precision, recall, and F1-score. The setup was reasonable, focusing on label efficiency, though the synthetic dataset’s limited diversity and lack of domain shift might inflate results. Real-world results align with expectations, but the small dataset size limits broader conclusions. The method’s superiority is evident, though potential overfitting to synthetic conditions remains a concern.
Further Thoughts
While the proposed framework demonstrates impressive label efficiency, I am intrigued by the potential limitations of synthetic data in fully capturing real-world single-photon imaging complexities, such as hardware-specific noise or environmental factors not modeled in the SIV module. Could discrepancies between synthetic and real data lead to biased uncertainty estimates, especially in deployment scenarios with unseen conditions? This connects to broader challenges in active learning for niche domains, where synthetic data is often a crutch due to scarce real data—similar issues are seen in medical imaging or autonomous driving under rare conditions. Additionally, the choice of ResNet-18, while practical, raises questions about whether architectures tailored to sparse, noisy data (e.g., inspired by graph neural networks for irregular data) could further enhance performance. Future work might explore hybrid models combining domain-specific priors with active learning. Lastly, the UIS metric’s robustness in edge cases (e.g., extreme noise levels) warrants deeper investigation, as over-reliance on synthetic variants might misguide sample selection if the simulation diverges from reality. These considerations could bridge this research to other low-data regimes in AI for science, where label efficiency is paramount.