MambaReID: A resource-efficient computer vision approach for generalized animal re-identification

Abstract

Successful wildlife conservation requires tracking and proper re-identification of individual animals in visibly challenged dense wild landscapes. The Royal Bengal Tiger of Bangladesh is an endangered species that acted as the major source of motivation to this research study and was eventually intended to develop advanced tools to carry out wildlife conservation activities. The initial phase concentrate on efficient animal re-identification via Mamba-based vision models. A comprehensive methodology was developed and evaluated using five deep learning models on a large-scale data set of 8,611 individual animals. The best performing model found, EfficientViM-M1, with only 14M parameters and 246M FLOPs, representing an 18× reduction in computational cost compared to Swin-T, achieved 83.13% Top-1 accuracy. Moreover, the metric learning approach ArcFace loss further improved individual animal separation, converging 4X faster than triplet loss and producing closer and well separated embedding clusters. These results confirm the effectiveness of our approach for resource constrained, real time wildlife monitoring study and lay the foundation for future work in this field.