AmphiLens wildlife detection checkpoints

This repository provides three domain-pretrained checkpoints for wildlife transfer learning, three later fine-tuned checkpoints for wildlife camera-trap imagery, and a machine-readable manifest. Training code and datasets are not included.

Project and research links

• AmphiLens application: github.com/joshsalako/amphilens
• Research project repository: github.com/joshsalako/wtl-detection

Research paper

Salako, J., Gordon, K., and Jeantet, L. Annotation-Efficient Object Detection of Endangered Western Leopard Toads in Camera Trap Imagery for Assessing Wildlife Tunnel Use.

Domain-pretrained checkpoints

These are the initial domain-pretraining checkpoints: they start from general pretrained detector weights and were trained on the combined wildlife sources iNaturalist, Open Images V7, California Small Animals, and Ohio Small Animals. They are the intermediate models used to initialize the later AmphiLens fine-tuning stage.

File	Architecture	Input size	Source labels	SHA-256	License status
yolo_domain.pt	YOLO26-M	640	Frog, Mouse, Snail	52b3a456db74b4ea6f310cf5ce7ea829d574854d79081de67ca8c9eb6d3c07c0	AGPL-3.0 (Ultralytics model)
rtdetr_domain.pt	RT-DETR-L	640	Frog, Mouse, Snail	ec7b6dc9180c2664b485f8f912bb8a4c8fe8f46168d2bd4300ab483d90be4b0f	AGPL-3.0 (Ultralytics model)
faster_rcnn_domain.pt	Faster R-CNN, ResNet-50 backbone	640	Frog, Mouse, Snail	97ed68343592d5502f389f3be70c0122ddc4a557e5d5160212159d5520847646	Unverified

Fine-tuned checkpoints

These are the later AmphiLens models fine-tuned for the three project labels.

File	Architecture	Phase 2 cycle	SHA-256	License status
yolo_clahe.pt	YOLO26-M	4	e4f8b3c4264b25482445506d0f495976f5af570ea50a4682a6d9ea247a8ea123	AGPL-3.0 (Ultralytics model)
rtdetr_clahe.pt	RT-DETR-L	3	91be4021b5f65017a3499651d45123475c7a3bc5776a313ebae18011732f42a9	AGPL-3.0 (Ultralytics model)
faster_rcnn_clahe.pt	Faster R-CNN, ResNet-50 backbone	5	ef69bc0691dcb765ad2af36d84ff3d536f4efe5e02c298b8c359b43d197c265e	Unverified

Labels

The domain-pretrained checkpoints use this class order:

1. Frog
2. Mouse
3. Snail

The fine-tuned checkpoints use this class order:

1. Other_Amphibian
2. Small_Mammal
3. Western_Leopard_Toad

Inference preprocessing

The fine-tuned checkpoints were trained with grayscale CLAHE preprocessing. The matching inference pipeline resizes each image with OpenCV linear interpolation, converts it to grayscale, applies CLAHE with clip limit 2.0 and an 8 × 8 tile grid, then replicates the result into three channels. Resized dimensions are rounded up to a multiple of 32.

• YOLO26-M and Faster R-CNN resize the shorter side to 640 pixels.
• RT-DETR resizes the longer side to a maximum of 640 pixels.
• YOLO inference uses imgsz=1152; RT-DETR uses imgsz=640.

In AmphiLens, the domain-pretrained entries use 640-pixel RGB inputs without additional grayscale or CLAHE preprocessing. The manifest records each checkpoint's architecture, class order, preprocessing profile, SHA-256 digest, and provenance. Downloads are pinned to the repository revision containing the selected checkpoint.

Use in AmphiLens

This repository is public and does not require login to download its checkpoints. To authenticate AmphiLens with your own Hugging Face account, run this on the machine running the app:

hf auth login

On first use, AmphiLens downloads the selected checkpoint to the Hugging Face cache and verifies its SHA-256 digest. For cloud training, the checkpoint is downloaded locally and uploaded through AmphiLens's checkpoint consent flow; the Hugging Face token is not included in that upload.

During inference, map each source class to a project class or choose Ignore; exact class-name matches are selected by default. Fine-tuning retains compatible feature weights and adapts the detector output head to the target classes.