CoolFace
Datasetpublic

lila-bc-community/channel-islands-camera-traps

Channel Islands Camera Traps — Detection Camera-trap images with bounding-box annotations from the Channel Islands, California, provided by The Nature Conservancy and distributed via LILA BC. Unofficial community mirror — not affiliated with or endorsed by LILA BC. This is an HF-native, viewer-ready object-detection build (parquet, dataset viewer renders boxes, trains in one command). For the full multi-dataset collection use the streaming loader… See the full description on the dataset page: https://huggingface.co/datasets/lila-bc-community/channel-islands-camera-traps.

sourceHugging Facecdla-permissive-1.0updated 4mo agoView on Hugging Face
0likes74downloads
Dataset Card

Channel Islands Camera Traps — Detection

Camera-trap images with bounding-box annotations from the Channel Islands, California, provided by The Nature Conservancy and distributed via LILA BC.

Unofficial community mirror — not affiliated with or endorsed by LILA BC. This is an HF-native, viewer-ready object-detection build (parquet, dataset viewer renders boxes, trains in one command). For the full multi-dataset collection use the streaming loader `society-ethics/lila_camera_traps`.

What this is

An example-sized subset: 12,000 bounding-box–annotated images sampled from the full dataset (~246k images, ~126k of which carry real animal boxes) using --boxed-only --max-images 12000. Empty frames (which LILA tags with an empty box) and human labels (LILA removes the images for privacy) are excluded — see Annotation cleaning.

  • —Images: 11,996
  • —Boxes: 14,843
  • —Classes: 5
ClassBoxes
fox8,236
bird4,638
rodent1,711
skunk253
other5

Structure

image     : Image
image_id  : string
width     : int32
height    : int32
objects   : { bbox: [x, y, w, h] (COCO xywh), category: ClassLabel, area: float32 }

Usage

python
from datasets import load_dataset

ds = load_dataset("lila-bc-community/channel-islands-camera-traps", split="train")

Train an object detector (one command)

bash
hf jobs uv run --flavor l4x1 \
  https://huggingface.co/datasets/hugging-face-vision-trainer/scripts/raw/main/object_detection_training.py \
  --dataset_name lila-bc-community/channel-islands-camera-traps \
  --model_name_or_path facebook/detr-resnet-50 \
  --output_dir channel-islands-detector --do_train --do_eval --push_to_hub

Annotation cleaning

Only genuine object boxes are kept: annotations with zero/negative-size boxes or non-object / privacy labels (empty, human, person, unidentified, …) are dropped. This is why the class set is the 5 animal classes rather than including empty/human.

Source, license & citation

The Nature Conservancy (2021): Channel Islands Camera Traps 1.0. The Nature Conservancy. Dataset.

For questions about the source dataset, contact Nathaniel Rindlaub (nathaniel.rindlaub@TNC.ORG).

See also