Class incremental learning for wildlife biodiversity monitoring in camera trap images

Abstract

Camera traps have been widely used for wildlife biodiversity monitoring, providing abundant ecological data. Manually classifying such abundant images is time-consuming and labor-intensive. Existing deep learning methods solve this problem for a fixed set of predefined wildlife species. The model trained on such sets cannot be applied to new wildlife species. Retraining models on new wildlife species can lead to catastrophic forgetting. Thus, in this work, we propose a class incremental learning method to identify new wildlife species. Our method employs a novel adaptive exemplar assignment (AEA) strategy with dynamic exemplar amounts to adapt to new species while alleviating the forgetting of old ones. Due to memory constraints, the data imbalance between limited exemplars and new species data can lead to class bias. We mitigate it by performing center vector retrieval (CVR) to classify samples in feature space and bypass the biased linear classifier. In addition, we propose two variants of CVR that incorporate the advantage of the linear classifier to further improve the performance. By using only 4% of old species data, our method achieves 77.09% accuracy at a low computational resource for recognition. Through extensive experiments and ablations, we demonstrate the superiority of our proposed approach over state-of-the-art methods. This method facilitates wildlife monitoring, biodiversity conservation, and ecological assessment.