Active Learning for Animal Re-Identification with Ambiguity-Aware Sampling

TL;DR

The paper addresses the challenge of animal re-identification under open-set conditions with limited annotations by introducing Ambiguity-Aware Sampling (AAS), which leverages disagreements between complementary clustering views to identify uncertain regions in the embedding space. It integrates a novel Non-Parametric, Plug-and-Play constrained clustering (NP3) to refine pseudo-labels using must-link and cannot-link feedback, enabling seamless use with existing unsupervised learning pipelines. Empirically, AAS achieves state-of-the-art performance across 13 wildlife datasets and two human Re-ID benchmarks using only a tiny annotation budget of $0.033 ext{ extpercent}$, with substantial gains in mAP and open-set metrics and statistically significant improvements. The work demonstrates practical, scalable improvements for biodiversity monitoring and wildlife conservation, highlighting structured uncertainty modeling as a powerful approach for real-world Re-ID systems.

Abstract

Animal Re-ID has recently gained substantial attention in the AI research community due to its high impact on biodiversity monitoring and unique research challenges arising from environmental factors. The subtle distinguishing patterns, handling new species and the inherent open-set nature make the problem even harder. To address these complexities, foundation models trained on labeled, large-scale and multi-species animal Re-ID datasets have recently been introduced to enable zero-shot Re-ID. However, our benchmarking reveals significant gaps in their zero-shot Re-ID performance for both known and unknown species. While this highlights the need for collecting labeled data in new domains, exhaustive annotation for Re-ID is laborious and requires domain expertise. Our analyses show that existing unsupervised (USL) and AL Re-ID methods underperform for animal Re-ID. To address these limitations, we introduce a novel AL Re-ID framework that leverages complementary clustering methods to uncover and target structurally ambiguous regions in the embedding space for mining pairs of samples that are both informative and broadly representative. Oracle feedback on these pairs, in the form of must-link and cannot-link constraints, facilitates a simple annotation interface, which naturally integrates with existing USL methods through our proposed constrained clustering refinement algorithm. Through extensive experiments, we demonstrate that, by utilizing only 0.033% of all annotations, our approach consistently outperforms existing foundational, USL and AL baselines. Specifically, we report an average improvement of 10.49%, 11.19% and 3.99% (mAP) on 13 wildlife datasets over foundational, USL and AL methods, respectively, while attaining state-of-the-art performance on each dataset. Furthermore, we also show an improvement of 11.09%, 8.2% and 2.06% for unknown individuals in an open-world setting.

Figures (3)

• Figure 1: An illustration of Ambiguity-Aware Sampling (AAS). (A.) Identifying Regions of Uncertainty. The disagreements between the clustering views $A$ and $B$ are used to identify regions of uncertainty, $S=\{S_1, S_2, S_3\}$, based on the transitive closure of partially overlapping clusters from $A$ and $B$. (B.) Resolving Over-Segmentation Errors. The medoid $\mathbf{m}_k$ is chosen to be the representative of an uncertain region $S_k$. For each medoid $\mathbf{m}_k$, we obtain its nearest $k_{max}$ neighboring medoids, $\mathbf{m}_{k'} ~\forall k' \neq k$, having similarity $>s_{min}$ to form most informative image pairs. (C.) Resolving Under-Segmentation Errors.$\tilde{I}_k$ is the set of all pairs that are in disagreement within $S_k$. Note that $\tilde{I}_k$ might contain redundant pairs that do not provide additional value if annotated. For instance, pairs $(x_1,x_6)$, $(x_2,x_6)$ and $(x_4,x_6)$ in the sub-figure represent the same ambiguity, i.e. whether the group of consistently clustered images $\{x_1, x_2, x_4\}$ should be clustered with $\{x_6\}$. Hence, to filter out redundant pairs, we define $\mathcal{P}_k^{cand}$ to be a set of candidate image pairs that contains the nearest inter-cluster neighbors for both the methods, which may or may not be inconsistent. Therefore, for each uncertain region $S_k$, we define the set of most informative and non-redundant inconsistent image pairs as $I_k=\tilde{I}_k ~\cap~ \mathcal{P}_k^{cand}$. Finally, the set of informative medoid pairs ($\mathcal{U}_{os}$) and non-redundant inconsistent pairs ($\mathcal{U}_{us}$) defines our sampling pool $\mathcal{U} = \{\mathcal{U}_{os} ~\cup~ \mathcal{U}_{us}\}$ from which the image pairs are sampled for annotation.
• Figure 2: Illustration of the proposed Non-Parametric, Plug-and-Play (NP3) algorithm for refining pseudo labels using pairwise constraints. Given an arbitrary initial clustering, NP3's first step is to satisfy all ML constraints by merging clusters that contain ML pairs. While this addresses ML constraints, it might lead to violations of CL constraints. To purify such impure clusters, our algorithm leverages a combination of graph coloring and Hungarian matching. Intuitively, all ML pairs inside an impure cluster are treated as single conceptual nodes for graph coloring, and any remaining CL pairs in the impure cluster are considered connected via edges. The graph is then colored, ensuring that all CL-constrained pairs are assigned different "colors" and thus separated. For any data points that remain unconstrained by either ML or CL pairs, a Hungarian matching is employed to associate them with their closest existing cluster, a strategy designed to minimize the overall change in the cluster structure.
• Figure 3: Performance comparison between base SpCL, existing AL Re-ID methods and AAS after every AL cycle. The numbers for AAS in (a) represent the total utilized budget. Since the utilized budget is approximately the same for all AL methods in (b) and (c), we remove them for clarity.