Table of Links

• Abstract and Introduction
• Related work
• The WildlifeDatasets toolkit
• MegaDescriptor – Methodology
• Ablation studies
• Performance evaluation
• Conclusion and References

5. Ablation studies

This section presents a set of ablation studies to empirically validate the design choices related to model distillation (i.e., selecting methods, architectures, and appropriate hyperparameters) while constructing the MegaDescriptor feature extractor, i.e., first-ever foundation model for animal re-identification. Furthermore, we provide both qualitative and quantitative performance evaluation comparing the newly proposed MegaDescriptor in a zero-shot setting with other methods, including SIFT, Superpoint, ImageNet, CLIP, and DINOv2.

5.1. Loss and backbone components

To determine the optimal metric learning loss function and backbone architecture configuration, we conducted an ablation study, comparing the performance (median accuracy) of ArcFace and Triplet loss with either a transformer- (Swin-B) or CNN-based backbone (EfficientNet-B3) on all available re-identification datasets. In most cases, the Swin-B with ArcFace combination maintains competitive or better performance than other variants. Overall, ArcFace and transformer-based backbone (Swin-B) performed better than Triplet and CNN backbone (EfficientNet-B3). First quantiles and top whiskers indicate that Triplet loss underperforms compared to ArcFace even with correctly set hyperparameters. The full comparison in the form of a box plot is provided in Figure 5.

5.2. Hyperparameter tunning

In order to overcome the performance sensitivity of metric learning approaches regarding hyperparameter selection and to select the generally optimal parameters, we have performed a comprehensive grid search strategy.

Following the results from the previous ablation, we evaluate how various hyperparameter settings affect the performance of a Swin-B backbone optimized with Arcface and Triplet losses. In the case of ArcFace, the best setting (i.e., lr = 0.001, m = 0.5, and s = 64) achieved a median performance of 87.3% with 25% and 75% quantiles of 49.2% and 96.4%, respectively. Interestingly, three settings underperformed by a significant margin, most likely due to unexpected divergence in the training[5]. The worst settings achieved a mean accuracy of 6.4%, 6.1%, and 4.0%. Compared to ArcFace, Triplet loss configurations showed higher performance on both 25% and 75% quantiles, indicating significant performance variability.

The outcomes of the study are visualized in Figure 6 as a boxplot, where each box consists of 29 values.

5.3. Metric learning vs. Local features

The results conducted over 29 datasets suggested that both metric learning approaches (Triplet and ArcFace) outperformed the local-feature-based methods on most datasets by a significant margin. The comparison of local-featurebased methods (SIFT and Superpoint) revealed that Superpoints are a better fit for animal re-identification, even though they are rarely used over SIFT descriptors in the literature. A detailed comparison is provided in Table 3. Note that the Giraffes dataset was labeled using local descriptors; hence, the performance is inflated and better than for metric learning.

The same experiment revealed that several datasets, e.g., AerialCattle2017, SMALST, MacaqueFaces, Giraffes, and AAUZebraFish, are solved or close to that point and should be omitted from development and benchmarking.

[5] These three settings were excluded from further evaluation and visualization for a more fair comparison.