Adaptive High-Frequency Transformer for Diverse Wildlife Re-Identification

Chenyue Li,Shuoyi Chen,Mang Ye

2024-10-25

Abstract:Wildlife ReID involves utilizing visual technology to identify specific individuals of wild animals in different scenarios, holding significant importance for wildlife conservation, ecological research, and environmental monitoring. Existing wildlife ReID methods are predominantly tailored to specific species, exhibiting limited applicability. Although some approaches leverage extensively studied person ReID techniques, they struggle to address the unique challenges posed by wildlife. Therefore, in this paper, we present a unified, multi-species general framework for wildlife ReID. Given that high-frequency information is a consistent representation of unique features in various species, significantly aiding in identifying contours and details such as fur textures, we propose the Adaptive High-Frequency Transformer model with the goal of enhancing high-frequency information learning. To mitigate the inevitable high-frequency interference in the wilderness environment, we introduce an object-aware high-frequency selection strategy to adaptively capture more valuable high-frequency components. Notably, we unify the experimental settings of multiple wildlife datasets for ReID, achieving superior performance over state-of-the-art ReID methods. In domain generalization scenarios, our approach demonstrates robust generalization to unknown species.

Computer Vision and Pattern Recognition,Artificial Intelligence

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the deficiencies of existing wildlife Re - ID (Re - identification) methods in terms of multi - species universality and adaptability. Specifically, most of the existing wildlife Re - ID methods are designed for specific species, resulting in the need to design methods separately for each species in practical applications, which severely limits universality and efficiency. Moreover, although some methods draw on mature pedestrian Re - ID techniques, these methods fail to fully address the unique challenges of wildlife, such as complex natural backgrounds and subtle differences between individuals.

To overcome these problems, the author proposes a unified re - identification framework applicable to multiple wildlife species - the Adaptive High - Frequency Transformer model. This model improves recognition accuracy by enhancing the learning of high - frequency information and introduces the following three strategies:

1. **Frequency - domain mixed data augmentation**: By mixing high - frequency information and original information at the frequency - domain level, the robustness of the model to environmental changes is enhanced.
2. **Object - aware dynamic selection**: Utilize the global attention mechanism to flexibly extract high - frequency regions related to the target, reducing the impact of environmental noise.
3. **Feature - balanced loss**: By introducing a loss function, balance the relationship between high - frequency features and original visual features to prevent over - emphasizing high - frequency details while ignoring the original information.

Through these strategies, this model can achieve better universality and adaptability among different species, significantly improving the accuracy and robustness of wildlife re - identification.

### Involved formulas

- **Frequency - domain mixing**:
\[
F'(I)=(1 - M_{\alpha})\cdot F_{h}(I)+M_{\alpha}\cdot F(I)
\]
where \(F_{h}(I)\) is the high - frequency component after high - pass filtering, and \(M_{\alpha}\) is a random matrix that controls the mixing ratio.

- **Feature - balanced loss**:
\[
L_{F}=\frac{1}{B}\sum_{b = 1}^{B}\frac{1}{Z}\sum_{z = 1}^{Z}\left\lVert f_{o}^{b,z}-f_{h}^{b,z}\right\rVert
\]
where \(f_{o}^{b,z}\) and \(f_{h}^{b,z}\) represent the original feature and high - frequency feature of the \(z\)-th token of the \(b\)-th input respectively.

- **Total loss function**:
\[
L_{\text{overall}}=L+\lambda L_{F}
\]
where \(L\) is the original ID loss and triplet loss, and \(\lambda\) is the weight of the feature - balanced loss.

Through these methods, this model can not only handle the re - identification tasks of multiple wildlife species, but also perform excellently in the domain generalization scenario and can identify unknown species.