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

TL;DR

Single-DGOD suffers from unseen, continuously evolving domain shifts that static augmentations struggle to model. LTFE addresses this by combining progressive Gaussian perturbations, temporal dependency modeling via LSTM, and liquid neural network–driven neural-ODE parameter evolution to simulate smooth feature trajectories from the source to latent target domains, reinforced by a temporal feature alignment module. The approach yields state-of-the-art generalization on Diverse Weather and Reality-to-Art benchmarks, with ablations demonstrating the complementary value of progressive perturbations, temporal dynamics, and semantic alignment. LTFE offers robust cross-domain detection with efficient inference, reducing reliance on target-domain data for deployment.

Abstract

In this paper, we focus on Single-Domain Generalized Object Detection (Single-DGOD), aiming to transfer a detector trained on one source domain to multiple unknown domains. Existing methods for Single-DGOD typically rely on discrete data augmentation or static perturbation methods to expand data diversity, thereby mitigating the lack of access to target domain data. However, in real-world scenarios such as changes in weather or lighting conditions, domain shifts often occur continuously and gradually. Discrete augmentations and static perturbations fail to effectively capture the dynamic variation of feature distributions, thereby limiting the model's ability to perceive fine-grained cross-domain differences. To this end, we propose a new method, Liquid Temporal Feature Evolution, which simulates the progressive evolution of features from the source domain to simulated latent distributions by incorporating temporal modeling and liquid neural network-driven parameter adjustment. Specifically, we introduce controllable Gaussian noise injection and multi-scale Gaussian blurring to simulate initial feature perturbations, followed by temporal modeling and a liquid parameter adjustment mechanism to generate adaptive modulation parameters, enabling a smooth and continuous adaptation across domains. By capturing progressive cross-domain feature evolution and dynamically regulating adaptation paths, our method bridges the source-unknown domain distribution gap, significantly boosting generalization and robustness to unseen shifts. Significant performance improvements on the Diverse Weather dataset and Real-to-Art benchmark demonstrate the superiority of our method. Our code is available at https://github.com/2490o/LTFE.

Figures (7)

• Figure 1: Liquid Temporal Feature Evolution (LTFE) for detecting the unknown-domain object. The core of LTFE lies in simulating the feature evolution trajectory from the source domain to simulated latent distributions, capturing continuous cross-domain feature evolution. As illustrated, LTFE facilitates a smooth transition of features from the source domain to the simulated latent distributions.
• Figure 2: Illustration of Liquid Temporal Feature Evolution. First, the first-layer feature map $F_0$ extracted by the backbone network is iteratively perturbed to generate the initial feature sequence $\{F_t\}_{t=1}^T$, simulating the feature transitions from the source domain to a latent domain. Next, LSTM is employed to model the generated feature sequence, capturing the spatiotemporal dependencies of feature evolution. Then, liquid parameter evolution is applied to dynamically adjust the original feature sequence, yielding the final evolved sequence $\{\hat{F}_t\}_{t=1}^T$, enabling a smooth and continuous transition between the source domain and the potential domain distributions. Finally, an alignment loss is used to further constrain the entire feature evolution process.
• Figure 3: Qualitative Results: Detection results under different weather conditions. The first and second rows display the results from G-NAS G-NAS and our method, respectively. To provide a more intuitive comparison, we highlight the objects missed or incorrectly detected by G-NAS G-NAS using red boxes, which are correctly identified by our method.
• Figure 4: Visualization analysis of our method: The first column displays the model's detection results. The second and third columns correspond to the channel activation visualizations of the initial feature $F_0$ and the evolved feature $\hat{F}_T$, respectively. The fourth and fifth columns display the heatmap visualizations of the initial feature $F_0$ and the evolved feature $\hat{F}_T$.
• Figure 5: Analysis of the time step $T$ in the Progressive Temporal Feature Evolution.