Facial landmark detection is a fundamental task in animal behavior analysis, animal welfare, and affective computing. However, existing approaches suffer from limited generalization due to sparse annotations, speciesspecific datasets, and morphological diversity across species. In this paper, we propose SuperAnimalFace, a unified pre-training framework designed to enhance wildlife facial landmark detection by leveraging diverse cross-species datasets. Central to our method is the construction of a panoramic facial landmark superset that enables semantic alignment across heterogeneous annotations. Additionally, we construct the WildLifeFLM dataset, the first largescale, multi-species benchmark with rich 25-point facial annotations. Our framework demonstrates strong zeroshot generalization results with several different backbones and high data efficiency in few-shot learning, achieving competitive performance with as little as 10% fine-tuning data. Experiments on a newly collected PetFace-45 validation set further prove that SuperAnimalFace provides a scalable and robust solution for the detection of animal facial landmarks. The proposed datasets are publicly available to the research community.
- Article type
- Year
- Co-author
Open Access
Issue
Open Access
Research Article
Issue
Tidal flat environments present unique challenges for Three Dimensional (3D) reconstruction and multimodal data matching due to their complex geological structures. Traditional methods for 3D modeling struggle with sparse feature distributions, leading to inaccuracies in both geometry and appearance. This paper presents a novel algorithm that combines neural implicit representations with multimodal data fusion to address these challenges. We introduce an attention-based strategy for multimodal feature alignment. It effectively aligns fine-grained optical image features with the structural and intensity information derived from Light Detection And Ranging (LiDAR) ambient images. Our method uniquely integrates self-distilled vision transformer with no labels v2 (namely DINOv2)-guided attention for fusing optical and LiDAR data. Additionally, we leverage Signed Distance Field (SDF) priors to explicitly separate geometry and appearance features, enabling precise surface reconstruction. Experimental results demonstrate that our algorithm outperforms the existing methods, particularly in reconstructing fine details and adapting to complex environmental conditions. Future work will explore incorporating point cloud data and temporal modeling to enhance robustness in highly dynamic tidal flat environments.
京公网安备11010802044758号