Single-Domain Generalized Object Detection (SDGOD) aims to achieve robust detection on unseen domains using only a single labeled source domain. Existing methods typically apply image- or feature-level perturbations independently, limiting their ability to model diverse domain shifts. To address this, we propose Fourier-based Dual-level Perturbation (FDP), which injects Gaussian perturbations into the amplitude spectra of both images and features in the Fourier domain. FDP consists of three modules: a Pseudo Domain Diversity (PDD) module that generates diverse pseudo-domains through spatial-frequency transformations, enriching image-level diversity. a Domain-Specific Spectral Perturbation (DSSP) module that introduces learnable low-frequency perturbations in feature space. and a Consistency Loss (CL) that encourages domain-invariant representation learning across perturbed views. Comparative experiments with ten mainstream SDGOD algorithms demonstrate that FDP achieves superior performance in single-domain generalized object detection tasks.

Facial age estimation aims to rank the face aging data by taking in the correlation among age categories. Conventional age estimation models are trained based on assumed high-quality training annotations in a totally-supervised manner. However, noisy data in a sparse distribution collected from unconstrained environment usually account for the corruption of produced gradients and ordinal relationships, which may fail to fairly describe the correlated face aging data. In this paper, we propose a meta-set learning (MSL) approach for exploiting the unfairness of face aging datasets, thus achieving unbiased age classification in unconstrained conditions. To address this, we elaborately create an unfairness filtration network under the meta-learning paradigm, which exceeds a reliable margin-reweighting initialization suffering from class variance, simultaneously exploiting the meta-reweighting intervention to minimize the training bias caused by class imbalance. Moreover, our proposed model leverages the learned instance-level margin between logits and develops a unimodal constrained logits loss, further surviving age regression models from unfairness. Experimental results on multiple in-the-wild datasets demonstrate that our proposed method achieves superior results compared to existing state-of-the-art methods.

In this paper, we introduce global compression learning (GCL) for finding reduced network architecture from a pre-trained network by removing both intra-layer and inter-layer structural redundancy. To accomplish this, we first derive architecture features from a binary representation of the network structure that effectively characterize the relationships between different layers. We then leverage reinforcement learning to iteratively compress the network via deep deterministic policy gradient based on the learned architecture features. To void extensive exploration of the huge space of network architectures, we bound feasible solutions within a small subspace by following a strict accuracy loss tolerance. Benchmarking tests show GCL outperforms the state-of-the-art models. On CIFAR-10 dataset, our model reduces 60.5% FLOPs and 93.3% parameters on VGG-16 without hurting the network accuracy, and yields a significantly compressed architecture for ResNet-110 by reductions of 71.92% FLOPs and 79.62% parameters with the cost of only 0.11% accuracy loss.