The same image can be described at multiple levels of detail (e.g., "a dog" vs. "a golden retriever sleeping on a couch" ), and humans naturally organize these descriptions into a hierarchy from coarse to fine. A practical retrieval system would better align with human recognition if respecting this semantic hierarchy during search. Hyperbolic geometry is well-suited to represent such hierarchical relations, but due to the geometry discrepancy, existing hyperbolic VLMs often demand costly retraining. We propose HALO (Hyperbolic Adaptation via Lifted Overlay), a lightweight hyperbolic retrieval framework that equips Euclidean-pretrained VLMs with hierarchy-aware search at low computational overhead, without compromising retrieval performance. We shift Lorentzian modeling to the embedding space where cross-modal similarity is computed, while keeping parameter updates fully Euclidean and LoRA-friendly. This design enables plug-and-play fine-tuning without entailment-style auxiliary losses used by prior hyperbolic VLMs. Empirically, HALO matches strong Euclidean pretrained baselines on retrieval performance. Across CLIP ViT-B/32, ViT-B/16, and ViT-L/14, our setting improves zero-shot recall by ~ 70-90% over pretrained hyperbolic (hierarchy-aware) baselines and by ~ 2%-6% over standard pre-trained CLIP baselines, while equipping euclidean CLIP with hierarchy-aware search.

In this paper, we introduce hierarchical action search. Starting from the observation that hierarchies are mostly ignored in the action literature, we retrieve not only individual actions but also relevant and related actions, given an action name or video example as input. We propose a hyperbolic action network, which is centered around a hyperbolic space shared by action hierarchies and videos. Our discriminative hyperbolic embedding projects actions on the shared space while jointly optimizing hypernym-hyponym relations between action pairs and a large margin separation between all actions. The projected actions serve as hyperbolic prototypes that we match with projected video representations. The result is a learned space where videos are positioned in entailment cones formed by different subtrees. To perform search in this space, we start from a query and increasingly enlarge its entailment cone to retrieve hierarchically relevant action videos. Experiments on three action datasets with new hierarchy annotations show the effectiveness of our approach for hierarchical action search by name and by video example, regardless of whether queried actions have been seen or not during training. Our implementation is available at https://github.com/Tenglon/hyperbolic_action.

Images classification is an essential field in remote sensing community. However, a variety of target shapes, as well as changing conditions during multiple time periods and different areas usually result in shifts in classification. This problem can affect classification results seriously. Although the affection is significant in remote sensing classification, very few people have considered this issue, and have solved it. In this paper, we introduce the associative domain adaptation (ADA) method to address this challenge. We apply this algorithm to two public remote sensing datasets. One is famous UC Merced dataset; another is NWPU-RESISC45 dataset which has a much more variance within the class. We then build a classification model by using UC Merced training images and labels as well as using training images from NWPU-RESISC45. This semi-supervised classification performance achieves an impressive test accuracy on the NWPU-RESISC45 test dataset.

Zero-shot learning (ZSL) aims to recognize unseen classes that are excluded from training classes. ZSL suffers from 1) Zero-shot bias (Z-Bias) --- model is biased towards seen classes because unseen data is inaccessible for training; 2) Zero-shot variance (Z-Variance) --- associating different images to same semantic embedding yields large associating error. To reduce Z-Bias, we propose a pseudo transfer mechanism, where we first synthesize the distribution of unseen data using semantic embeddings, then we minimize the mismatch between the seen distribution and the synthesized unseen distribution. To reduce Z-Variance, we implicitly corrupted one semantic embedding multiple times to generate image-wise semantic vectors, with which our model learn robust classifiers. Lastly, we integrate our Z-Bias and Z-variance reduction techniques with a linear ZSL model to show its usefulness. Our proposed model successfully overcomes the Z-bias and Z-variance problems. Extensive experiments on five benchmark datasets including ImageNet-1K demonstrate that our model outperforms the state-of-the-art methods with fast training.

With the development of machine learning, many researchers have used machine learning models for building extraction in high resolution remote sensing images. Especially for the recently proposed deep learning models, it has been widely used for building detection in the urban monitoring. In this study, the performances of images segmentation for building extraction based on Fully Convolutional Networks (FCN) model and shallow models are qualitatively and quantitatively compared. Firstly, the public aerial dataset of Massachusetts building dataset[1] are preprocessed to extract features. Then, we trained shallow models and deep model on Massachusetts building dataset. Moreover, the trained FCN model and shallow models are used to extract the building on the same image. Finally, we compared performances between shallow models and deep model. It is found that FCN gives the highest recall 0.63, precision 0.62, and F-measure rate 0.63. The qualitative and quantitative analysis of the building extraction results fully demonstrates that FCN gives the best performance compared with traditional shallow models.