Abstract:
Visual wildlife monitoring of animals requires detection for species-level categorization and re-identification (Re-ID) for population estimation of an individual species. Traditionally, the monitoring is done via GPS collars which are invasive, but the advent of camera traps has given a convenient, non-invasive and inexpensive alternate method for monitoring of wild animals. This camera-trap image data can be used with AI-based algorithms for detecting animal presence, species-level categorization, as well as individual identification or animal biometrics for certain species. To this end, we have developed the Deep Learning (DL) based species categorization module for the Camera Trap Data Repository and Analysis Tool (CaTRAT), which was used by theWildlife Institute of India (WII) for processing the camera trap images during the All India Tiger Estimation 2022. Beyond species-level segregation, deep learning approaches have also shown good performance for re-identification tasks. However, these methods often fall short, when encountered with fine grained patterned species like tigers and leopards, both in terms of performance as well as interpretability. This limits their usability by conservation officials and practitioners. In this work, we propose an end-to-end network to learn feature representations, keypoints, and their descriptors. The keypoints enable the model to: a) learn better discriminative feature representations and b) focus on salient regions (patterns) of the image. It is important to note that while training, we don’t have groundtruth keypoint and descriptor annotations but only the label information. A pre-trained, DenseNet model is fine-tuned by a classification cross-entropy loss regularized by a pairwise Jensen-Shannon divergence. Further, feature map normalization regularizes the descriptor loss. The fusion of the keypoint attention feature map in the network helps focus on regions important for animal biometrics. We then evaluate the efficacy of our model on two datasets of patterned species, namely Amur Tigers and Leopards, under different biometric evaluation protocols: mAP, top-1, top-5, closed-set identification, open-set identification, and verification.
