Implicit Non-Causal Factors are Out via Dataset Splitting for Domain Generalization Object Detection

TL;DR

Open-world domain generalization for object detection is challenged by non-causal factors hidden in domain-invariant features. The authors introduce GB-DAL, which generates dense domain labels via Prototype-based Granular Ball Splitting (PGBS), and SNF, which simulates non-causal perturbations for data augmentation, forming a causal-learning framework that improves generalization. Evaluations across six cross-domain DGOD benchmarks show GB-DAL consistently outperforms standard DAL and DG baselines, with SNF providing additive gains and robustness to adversarial and natural noise. The approach also transfers to image classification (e.g., PACS), underscoring its broad applicability and practical impact in open-world vision tasks.

Abstract

Open world object detection faces a significant challenge in domain-invariant representation, i.e., implicit non-causal factors. Most domain generalization (DG) methods based on domain adversarial learning (DAL) pay much attention to learn domain-invariant information, but often overlook the potential non-causal factors. We unveil two critical causes: 1) The domain discriminator-based DAL method is subject to the extremely sparse domain label, i.e., assigning only one domain label to each dataset, thus can only associate explicit non-causal factor, which is incredibly limited. 2) The non-causal factors, induced by unidentified data bias, are excessively implicit and cannot be solely discerned by conventional DAL paradigm. Based on these key findings, inspired by the Granular-Ball perspective, we propose an improved DAL method, i.e., GB-DAL. The proposed GB-DAL utilizes Prototype-based Granular Ball Splitting (PGBS) module to generate more dense domains from limited datasets, akin to more fine-grained granular balls, indicating more potential non-causal factors. Inspired by adversarial perturbations akin to non-causal factors, we propose a Simulated Non-causal Factors (SNF) module as a means of data augmentation to reduce the implicitness of non-causal factors, and facilitate the training of GB-DAL. Comparative experiments on numerous benchmarks demonstrate that our method achieves better generalization performance in novel circumstances.

Figures (8)

• Figure 1: Visual representations of inter- and intra-dataset non-causal factors. Road condition that differ in the two datasets are regarded as inter-dataset non-causal factors, while vehicle color that reflect data bias within the dataset are considered as intra-dataset non-causal factors.
• Figure 2: Comparison of the alignment effects of DAL based on sparse and dense domain labels. (a) With sparse domain labels (i.e., dataset-level), DAL eliminates the inter-dataset non-causal factors (represented by color), but ignoring intra-dataset non-causal factors (represented by filling pattern). (b) With denser domain labels (i.e., datasets splits), intra-dataset non-causal factors can also be filtered out.
• Figure 3: Results of the domain label assignment of the PGBS module on cars under spurious correlations in NICO he2021towards.
• Figure 4: Diagram of the proposed domain generalized object detector, featuring two novel components: GB-DAL and SNF. SNF employs adversarial perturbations to simulate and expose hidden non-causal factors. In GB-DAL, the PGBS module (Prototype-based Granular Ball Splitting) computes dense domain labels via datasets split for domain adversarial learning, facilitating the identification and filtering of additional non-causal factors. GRL means the Gradient Reversal Layer.
• Figure 5: Illustration of the Prototype-based Granular Ball Splitting (PGBS) module, which conducts datasets split.