An AI toad detector joins a Cape Town conservation team

Nature Connect has integrated Joshua Salako’s AmphiLens detector into its weekly conservation work. Its 26 October report describes a staff portal built by Llewellyn Williams, where people confirm flagged images.[1]

The work begins with a road between two wetlands. Peninsula Road separates Rondevlei and Zeekoevlei in Cape Town’s False Bay Nature Reserve. Endangered Western Leopard Toads move towards breeding wetlands at night between July and October, bringing them into contact with traffic. Six underpasses, constructed between 2021 and 2023, provide a route beneath the road; 800 metres of fencing guide the animals towards the tunnels.[2]

For the people looking after that route, monitoring has meant much more than installing cameras. Nature Connect’s account describes volunteers, students and staff spending 467 hours on nocturnal patrols over 96 nights in 2023–2025. They recorded amphibian activity, road deaths, breaches in the barriers and environmental conditions. The resulting photographs added another source of evidence, and another substantial task.[2]

A research problem close to home

Salako approached that task as an AI for Science master’s student at the African Institute for Mathematical Sciences in South Africa, supervised by Lorène Jeantet. The institute’s newsletter describes a collaboration with Nature Connect using camera images collected nearby. By its July newsletter, posted in August, the work had been accepted for the Computer Vision for Ecology workshop at ECCV 2026.[3]

The research began with 20 labelled images. Its published code describes a pipeline that improves contrast in dark frames, adapts models to wildlife imagery and asks people to label selected examples in successive rounds. Four detector families were compared: YOLO, RT-DETR, Faster R-CNN and MegaDetector. Faster R-CNN gave the strongest Western Leopard Toad results in the authors’ evaluation.[4]

That last point makes the method more interesting than a simple search through an archive. Labelling is itself work: someone has to identify the animal and mark its location. Active learning directs that effort towards images expected to improve the model. The research implementation combines uncertainty-based selection with a second selection method intended to provide useful visual variety.[4]

The application’s training guide carries the idea into a repeatable process. A project keeps labelled datasets as fixed snapshots, supports separate training, validation and test material, and returns selected images to the annotation workflow. Once people have supplied the next labels, training can continue. The stored snapshots give the team a record of the material behind each cycle.[5]

A small number of sightings in a large archive

Nature Connect is still checking individual identities and crossings. The portal proposes matches between back patterns; a person confirms each match.[1]

The cameras photograph at 15-second intervals, so a fast-moving toad can pass between frames. Some journeys may therefore go unrecorded.[1]

A tool other projects can adapt

AmphiLens is available as code with installation instructions and a browser interface that runs locally. Its documented workflow starts with an image folder, then detection and review, with optional labelling and further training. Setups cover ordinary CPU processing, compatible Nvidia GPUs and Apple Silicon Macs, as well as cloud jobs. Using it requires installing the software; it is not simply a public upload page.[7]

The published model collection also exposes the stages behind the detector. Initial wildlife models were trained using material from iNaturalist, Open Images V7 and two small-animal datasets. Later models were adapted to three project categories: Western Leopard Toad, other amphibians and small mammals. Their model cards specify the image preparation used during training, including the contrast treatment that must also be applied when those models analyse new photographs.[6]

For another conservation team, that offers a starting point with a visible route to adaptation. The app keeps project results separate from original photographs and supports further training for the species a team wants to study. Its default local workflow also lets a project work with its own image collection before choosing whether to use remote annotation or cloud training.[7]

At tunnel 4, one confirmed crossing follows toad T2026-7 from Rondevlei at 00:38 on 21 August to Zeekoevlei eight minutes later, by the camera clocks. Among the empty frames is that short, documented journey.[1]

Report an error or suggest a correction