Lion preservation with computer vision
The Lion Identification Network of Collaborators - LINC - monitors lions to help with their preservation across African territory. This community of conservationists and researchers in conjunction with IEF R&D created an open-source platform to track Panthera Leo in order to maintain the genetic viability of the species and create better policy decisions to protect the last remaining lions.
They do so by manually identifying whisker patterns taken from photographs, and preset grids where each pattern corresponds to one specimen. Another solution was to place a short-term GPS-tracking collar on each lion.
Both of the options are time-consuming, and collaring the lions is especially invasive for the animal.
The tool consists of computer vision and pattern recognition algorithms that can automatically perform two different methods for lion identification: face and whisker id.
To achieve this, the community uploads the photographs into the system, which then performs a database match to identify each specific lion. This reduces the time spent and team size needed, allowing conservationists to use large data sets in their work viably.
Organization type: NGO
Helping wild animals conservation with machine learning.
A computer vision approach proved to be successful in animal tracking. Get inspired by other AI for social good applications.
Develop a computer vision automated solution and pattern recognition software in 3 stages:
Lion feature object detection
With a metric learning approach, the face images are mapped to a vectorial space, where two images from the same lion are very close, and pictures from different lions are distant.
A network is trained as an image classifier to achieve this mapping, where each class represents a different entity (in this case, a different lion). The classifier layers are removed, and the embeddings generated by an intermediate layer is used as our image representation.
Transfer learning over a pre-trained ResNet50 network is applied, and then the last three layers are removed, getting an output of size 512. All the experiments from this phase are carried out using FastAI, a library built on top of PyTorch.
All images are labeled with the lion's identifier, and given a photo of a specimen's whisker area, the system returns the lions from the database with the highest probability of being the one portrayed in the picture.
Phase one object detector is used for detecting single whisker spots, this generates a point cloud from each of the whisker images. Since very few images per lion are available at this dataset, experiments with point cloud algorithms were carried out.
A match between two of these point clouds is performed by applying Coherent Point Drift, and then a point cloud distance is used to determine how good is the pairing. Then the lions with the smallest matching distance are the ones with the highest probability of being the lion we are looking for.
Tryolabs was a great partner in developing the new algorithms for this uniquely challenging problem set. The end result of this work has been a targeted solution that can run in a resource constrained environment with or without network connection.
This work shows the flexibility and an openness of Tryolabs to out of the box thinking essential to developing not just a powerful algorithm but one that can work on all levels of the conservationist work practice.
Overall it has been a great working experience and IEF R&D and the conservation team look forward to working in the future with Tryolabs on this unique and important conservation challenge.