Simulating Distribution Dynamics: Liquid Temporal Feature Evolution for Single-Domain Generalized Object Detection

Zihao Zhang, Yang Li, Aming Wu, Yahong HanarXiv 2025 · DOI · arXiv

Abstract

Single-Domain Generalized Object Detection (Single-DGOD) aims to train models on a single source domain that can generalize to diverse, unseen target environments. Traditional methods often rely on static augmentations that fail to capture the continuous nature of real-world domain shifts, such as changing weather or lighting. This paper introduces Liquid Temporal Feature Evolution (LTFE), a framework that models domain shifts as continuous trajectories using Neural Ordinary Differential Equations (ODEs). By simulating a "flow" of feature degradation and employing liquid parameter evolution, LTFE achieves state-of-the-art performance across challenging scenarios like rainy nights and foggy days, significantly closing the distribution gap in unseen domains.

Introduction

In computer vision, the "Single-Domain Generalized Object Detection" (Single-DGOD) task presents a significant challenge: training a model on one specific environment (e.g., sunny daytime) and ensuring it remains robust in completely different, unseen conditions (e.g., rainy nights or foggy mornings). Most existing approaches treat these shifts as discrete jumps, using static data augmentation or text-based prompts to approximate new domains. However, environmental changes in the real world are inherently continuous and gradual.

Conceptual overview of feature evolution across different domains in the feature space.

The core insight of this work is to treat domain shifts as dynamic trajectories rather than static snapshots. As shown in Figure 1, the proposed Liquid Temporal Feature Evolution (LTFE) framework models the transition between domains as a continuous path in the feature space. By perceiving these fine-grained, progressive differences, the model can better align its internal representations with the reality of new environments.

Methodology

The LTFE architecture is built upon three primary pillars designed to simulate and adapt to continuous domain shifts.

The LTFE framework architecture, featuring Progressive Temporal Feature Evolution and Neural ODE-based parameter generation.

1. Progressive Temporal Feature Evolution (PTFE)

Instead of applying random, one-off augmentations, PTFE simulates a "degradation path." It applies a sequence of multi-scale Gaussian blurring and controllable Gaussian noise to the initial features ($F_0$). This process creates a temporal sequence of features ($F_1, \dots, F_T$) that represent a transition from clear to increasingly degraded states, mimicking real-world visibility loss.

2. Temporal Dependency and Liquid Parameter Evolution

To maintain logical consistency across the simulated degradation, an LSTM network tracks the evolution of features over time. The "liquid" aspect of the model comes from Liquid Parameter Evolution (LPE), which utilizes Neural Ordinary Differential Equations (Neural ODEs). Unlike standard fixed convolution kernels, these liquid parameters are dynamically generated along a continuous path. This allows the model's internal representations to "flow" and adapt smoothly to the target domain's specific distribution.

3. Temporal Feature Alignment Module (TFAM)

To prevent the loss of semantic information during adaptation, the TFAM ensures that the model retains its ability to identify objects (e.g., cars, pedestrians). It employs specific loss functions to maintain intra-class consistency and inter-class separability, keeping the semantic meaning of features intact even as they evolve to match new environmental conditions.

Experiments and Results

The LTFE framework was evaluated on two major benchmarks: the Diverse Weather Dataset (generalizing from Sunny to Night, Rainy, and Foggy conditions) and the Real-to-Art Benchmark (generalizing from real photos to artistic styles like Clipart and Watercolor).

Key Findings

  • Superior Robustness: LTFE consistently outperformed state-of-the-art (SOTA) methods. On the Diverse Weather dataset, it achieved a mean Average Precision (mAP) of 41.2%, surpassing previous leaders like SECOT (40.6%) and G-NAS (36.4%).
  • Extreme Scenario Performance: The model showed its greatest strength in the most difficult scenarios. In "Night Rainy" conditions, LTFE improved upon the baseline by nearly 11% mAP.
  • Versatility: Despite being designed for continuous shifts, the model also excelled at "abrupt" artistic shifts, beating the SECOT method by 4.7% on the Clipart dataset.
  • Efficiency: LTFE maintains a practical inference speed of 6.4 FPS on a ResNet-101 backbone, balancing the complexity of ODE solvers with real-time deployment needs.

Conclusion

By shifting the perspective from static domain adaptation to continuous trajectory modeling, Liquid Temporal Feature Evolution provides a robust solution for Single-DGOD. The use of Neural ODEs allows for a level of flexibility that standard networks lack, making the model uniquely suited for the fluctuating conditions of the real world. Future work will look into expanding this "liquid" approach to multi-source domain training and incorporating more complex environmental noise simulations beyond Gaussian models.

Summary updated

BibTeX
@misc{zhang2025simulating,
  author = {Zihao Zhang and Yang Li and Aming Wu and Yahong Han},
  title = {{Simulating Distribution Dynamics: Liquid Temporal Feature Evolution for Single-Domain Generalized Object Detection}},
  year = {2025},
  eprint = {2511.09909},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.09909},
}
Figure