ABSTRACT
Continuous Sign Language Recognition (CSLR) is a task which converts a sign language video into a gloss sequence. The existing deep learning based sign language recognition methods usually rely on large-scale training data and rich supervised information. However, current sign language datasets are limited, and they are only annotated at sentence-level rather than frame-level. Inadequate supervision of sign language data poses a serious challenge for sign language recognition, which may result in insufficient training of sign language recognition models. To address above problems, we propose a cross-modal knowledge distillation method for continuous sign language recognition, which contains two teacher models and one student model. One of the teacher models is the Sign2Text dialogue teacher model, which takes a sign language video and a dialogue sentence as input and outputs the sign language recognition result. The other teacher model is the Text2Gloss translation teacher model, which targets to translate a text sentence into a gloss sequence. Both teacher models can provide information-rich soft labels to assist the training of the student model, which is a general sign language recognition model. We conduct extensive experiments on multiple commonly used sign language datasets, i.e., PHOENIX 2014T, CSL-Daily and QSL, the results show that the proposed cross-modal knowledge distillation method can effectively improve the sign language recognition accuracy by transferring multi-modal information from teacher models to the student model. Code is available at https://github.com/glq-1992/cross-modal-knowledge-distillation_new.
Subject(s)
Deep Learning , Sign Language , Humans , Neural Networks, Computer , Distillation/methodsABSTRACT
Great efforts have been made to improve the soil and water conservation capacity by restoring plant communities in different climatic and land-use types. However, how to select suitable species from local species pools that not only adapt to different site environments, but also achieve certain soil and water conservation capacities is a great challenge in vegetation restoration for practitioners and scientists. So far, little attention has been paid to plant functional response and effect traits related to environment resource and ecosystem functions. In this study, together with soil properties and ecohydrological functions, we measured the seven plant functional traits for the most common species in different restoration communities in a subtropical mountain ecosystem. Multivariate optimization analyses were performed to identify the functional effect types and functional response types based on specific plant traits. We found that the community-weighted means of traits differed significantly among the four community types, and the plant functional traits were strongly linked with soil physicochemical properties and ecohydrological functions. Based on three optimal effect traits (specific leaf area, leaf size, and specific root length) and two response traits (specific leaf area and leaf nitrogen concentration), seven functional effect types in relation to the soil and water conservation capacity (interception of canopy and stemflow, maximum water-holding capacity of litter, maximum water-holding capacity of soil, soil surface runoff, and soil erosion) and two plant functional response types to soil physicochemical properties were identified. The redundancy analysis showed that the sum of all canonical eigenvalues only accounted for 21.6% of the variation in functional response types, which suggests that community effects on soil and water conservation cannot explain the overall structure of community responses related to soil resources. The eight overlapping species between the plant functional response types and functional effect types were ultimately selected as the key species for vegetation restoration. Based on the above results, we offer an ecological basis for choosing the appropriate species based on functional traits, which may be very helpful for practitioners involved in ecological restoration and management.
ABSTRACT
Functional characteristics are increasingly used to evaluate the success of different vegetation restoration. Community functional diversity (FD) and the community-weighted mean (CWM), as two main complementary components, are closely linked to site environment and ecosystem functions. However, the patterns and driving factors of functional characteristics are still not clear in different vegetation restoration types. Here, four community restoration types (secondary shrubland, SL; Pinus yunnanensis forest, PF; mixed needle-broad-leaved forest, MF; natural secondary forest, NSF) were selected to investigate species diversity, FD, CWM, and soil physicochemical properties. The relative effects of species diversity and soil abiotic features on variation in functional characteristics were then evaluated. We found that different restoration communities altered most community structures and functional properties in terms of species diversity, FD, and CWM. CWM values and FD in different communities presented different distribution patterns depending on certain traits and parameters. Significant correlations between functional traits were found at the species and community scales, suggesting a potential covariation between these selected traits in communities. The results of redundancy analysis and variation partitioning showed that most of the variation in functional characteristics, especially CWM, was explained by soil moisture and available phosphorus, indicating that habitat filters regulate the functional characteristics of plant communities mainly by changing the dominant species composition and functional traits of species. Therefore, the selection of restoration species adapted to low soil moisture and available phosphorus and the construction of communities based on selected species as the dominant species can effectively drive community assembly and ecosystem functions in the vegetation restoration process.
ABSTRACT
Influential node identification plays a significant role in understanding network structure and functions. Here we propose a general method for detecting influential nodes in a graph-traversal framework. We evaluate the influence of each node by constructing a breadth-first search (BFS) tree in which the target node is the root node. From the BFS tree, we generate a curve in which the x axis is the level number and the y axis is the cumulative scores of all nodes visited so far. We use the area under the curve value as the final influence score of the target node. Experimental results on various networks across different domains demonstrate that our method can be significantly superior to widely used centrality measures on the task of influential node detection.