SAN: Selective Alignment Network for Cross-Domain Pedestrian Detection

Cross-domain pedestrian detection via feature alignment and image quality assessment

Datasets collected under different sensors, viewpoints, or weather conditions cause different domains. Models trained on domain A applied to tasks of domain B result in low performance. To overcome the domain shift, we propose an unsupervised pedestrian detection method that utilizes CycleGAN to establish an intermediate domain and transform a large gap domain-shift problem into two feature alignment subtasks with small gaps. The intermediate domain trained with labels from domain A, after two rounds of feature alignment using adversarial learning, can facilitate effective detection in domain B. To further enhance the training quality of intermediate domain models, Image Quality Assessment (IQA) is incorporated. The experimental results evaluated on Citypersons, KITTI, and BDD100K show that MR of 24.58%, 33.66%, 28.27%, and 28.25% were achieved in four cross-domain scenarios. Compared with typical pedestrian detection models, our proposed method can better overcome the domain-shift problem and achieve competitive results.

Source-Guided Target Feature Reconstruction for Cross-Domain Classification and Detection

Existing cross-domain classification and detection methods usually apply a consistency constraint between the target sample and its self-augmentation for unsupervised learning without considering the essential source knowledge. In this paper, we propose a Source-guided Target Feature Reconstruction (STFR) module for cross-domain visual tasks, which applies source visual words to reconstruct the target features. Since the reconstructed target features contain the source knowledge, they can be treated as a bridge to connect the source and target domains. Therefore, using them for consistency learning can enhance the target representation and reduce the domain bias. Technically, source visual words are selected and updated according to the source feature distribution, and applied to reconstruct the given target feature via a weighted combination strategy. After that, consistency constraints are built between the reconstructed and original target features for domain alignment. Furthermore, STFR is connected with the optimal transportation algorithm theoretically, which explains the rationality of the proposed module. Extensive experiments on nine benchmarks and two cross-domain visual tasks prove the effectiveness of the proposed STFR module, e.g., 1) cross-domain image classification: obtaining average accuracy of 91.0%, 73.9%, and 87.4% on Office-31, Office-Home, and VisDA-2017, respectively; 2) cross-domain object detection: obtaining mAP of 44.50% on Cityscapes → Foggy Cityscapes, AP on car of 78.10% on Cityscapes → KITTI, MR -2 of 8.63%, 12.27%, 22.10%, and 40.58% on COCOPersons → Caltech, CityPersons → Caltech, COCOPersons → CityPersons, and Caltech → CityPersons, respectively.

Probability-based Graph Embedding Cross-domain and Class Discriminative Feature Learning for Domain Adaptation

Feature-based domain adaptation methods project samples from different domains into the same feature space and try to align the distribution of two domains to learn an effective transferable model. The vital problem is how to find a proper way to reduce the domain shift and improve the discriminability of features. To address the above issues, we propose a unified Probability-based Graph embedding Cross-domain and class Discriminative feature learning framework for unsupervised domain adaptation (PGCD). Specifically, we propose novel graph embedding structures to be the class discriminative transfer feature learning item and cross-domain alignment item, which can make the same-category samples compact in each domain, and fully align the local and global geometric structure across domains. Besides, two theoretical analyses are given to prove the interpretability of the proposed graph structures, which can further describe the relationships between samples to samples in single-domain and cross-domain transfer feature learning scenarios. Moreover, we adopt novel weight strategies via probability information to generate robust centroids in each proposed item to enhance the accuracy of transfer feature learning and reduce the error accumulation. Compared with the advanced approaches by comprehensive experiments, the promising performance on the benchmark datasets verify the effectiveness of the proposed model.