Using higher-resolution feature maps in the network is an effective approach for detecting small objects. However, high-resolution feature maps face the challenge of lacking semantic information. This has led previous methods to rely on downsampling feature maps, applying large-kernel convolution layers, and then upsampling the feature maps to obtain semantic information. However, these methods have certain limitations: first, large kernel convolutions in deeper layers typically provide significant global semantic information, but our experiments reveal that such prominent semantic information introduces background smear, which in turn leads to overfitting. Second, deep features often contain substantial redundant information, and the features of small objects are either minimal or have disappeared, which causes a degradation in detection performance when directly relying on deep features. To address these issues, we propose a high-resolution network based on local contextual semantics (HR-SemNet). The network is built on the proposed high-resolution backbone (HRB), which replaces the traditional backbone-FPN architecture by focusing all computational resources of large kernel convolutions on high-resolution feature layers to capture clearer features of small objects. Additionally, a local context semantic module (LCSM) is employed to extract semantic information from the background, confining the semantic extraction to a local window to avoid interference from large-scale backgrounds and objects. HR-SemNet decouples small object semantics from contextual semantics, with HRB and LCSM independently extracting these features. Extensive experiments and comprehensive evaluations on the VisDrone, AI-TOD, and TinyPerson datasets validate the effectiveness of the method. On the VisDrone dataset, which contains a large number of small objects, HR-SemNet improves the mean average precision (mAP) by 4.6%, reduces the computational cost (GFLOPs) by 49.9%, and decreases the parameter count by 94.9%.

Unsupervised domain adaptation, which transfers knowledge from the source domain to the target domain, has still been a challenging problem. However, previous domain adaptation methods typically minimize the domain discrepancy by using the pseudo target labels. Since the pseudo labels can be noisy, which may cause misalignment and unsatisfying adaptation performance. To address the above challenges, we propose an information maximization adaptation network with label distribution priors. We revisit feature alignment in unsupervised domain adaptation from the perspective of distribution alignment, and find that learning discriminant feature representation requires to minimizing distribution discrepancy and maximizing source mutual information between the outputs of the classifier and feature representations. Due to domain shift, maximizing target mutual information may align features to incorrect class directly. We propose a weighted target mutual information by re-weighting the estimated mutual information via the mean prediction confidence in mini-batch, which can eliminate the negative impact of inaccurate estimation. In addition, we introduce a regularization term of label priors distribution to encourage the similarity to the real label distribution. Extensive experimental results on three benchmark datasets show that our proposed method can achieve remarkable results compared with previous methods.

Deep learning has achieved remarkable progress for visual recognition on balanced data sets but still performs poorly on real-world long-tailed data distribution. The existing methods mainly decouple the problem into the two-stage decoupling training, that is, representation learning and classifier training, or multistage training based on knowledge distillation, thus resulting in huge training steps and extra computation cost. In this paper, we propose a conceptually simple yet effective One-stage Long-tailed Self-Distillation framework, called OLSD, which simultaneously takes representation learning and classifier training into one-stage training. For representation learning, we take two different sampling distributions and mixup them to input them into two branches, where the collaborative consistency loss is introduced to train network consistency, and we theoretically show that the proposed mixup naturally generates a tail-majority distribution mixup. For classifier training, we introduce balanced self-distillation guided knowledge transfer to improve generalization performance, where we theoretically show that proposed knowledge transfer implicitly minimizes not only cross-entropy but also KL divergence between head-to-tail and tail-to-head. Extensive experiments on long-tailed CIFAR10/100, ImageNet-LT and multilabel long-tailed VOC-LT demonstrate the proposed method's effectiveness.

Recent advances in deep neural network have achieved remarkable successes in various computer vision tasks. However, deep neural network with millions of parameters may suffer from poor generalization due to overfitting. To improve the generalization performance, many methods have been proposed such as data augmentation, label smoothing and knowledge distillation. In this paper, we extend self-knowledge distillation to enhance the model generalization performance without incurring extra computation cost, called EWR-KD, which is a simple yet effective method to progressively distill knowledge from the model itself. Concretely, it consists of two components: 1) the self-distillation scheme that progressively softens the learning targets by using the past model prediction; 2) the sample-reweighting scheme that dynamically decides the trust degree to transfer more informative knowledge by introducing uncertainty estimation. With the two components, EWR-KD is robust to both corrupt noises and adversarial noises, and can be easily combined with current advanced regularization techniques. We theoretically show that EWR-KD minimizes cross-entropy by adding an epoch-wise regularization, which measures the difference between the past model prediction at ($t-1$)-th epoch and the current prediction at$t$-th epoch. Finally, Extensive experimental results on clean datasets and noisy datasets empirically demonstrate that EWR-KD not only improves the performance of the state-of-the-art baseline but also yields well calibration.