Snapshot Safari SER (Serengeti) — Sampled Classroom Subset v1.0

Summary

A curated, realistically distributed subset of the Snapshot Safari 2024 Expansion SER (Serengeti National Park) camera trap dataset, prepared for use in the CAS Deep Learning — Computer Vision course exercises. Unlike ser_balanced, the class distribution reflects real-world Serengeti encounter rates: wildebeest and zebra dominate, while rarer species (impala, warthog, gazelle grants) appear infrequently.

Archive	Images	Description
`ser_sampled.tar.gz`	4 999	Realistic distribution, full frames
`ser_sampled_cropped.tar.gz`	4 949	Realistic distribution, MD-cropped

Source

Dataset: Snapshot Safari 2024 Expansion — SER (Serengeti) subset
URL: https://lila.science/datasets/snapshot-safari-2024-expansion/
License: Community Data License Agreement — Permissive variant 1.0
Attribution: Snapshot Safari / University of Minnesota Lion Center

MegaDetector

Pre-computed MegaDetector v1000-redwood RDE-filtered results from LILA Science: snapshot-safari-2024-expansion-SER-subset-v1000.0.0-redwood_detections.threshold.filtered.json.zip

Used to filter frames (conf ≥ 0.8) and select the best frame per sequence. The _cropped variant additionally crops each image to the primary detected animal bounding box (10% padding).

Species

buffalo, elephant, empty, gazellegrants, gazellethomsons, hartebeest, impala, warthog, wildebeestblue, zebraplains

Statistics — `ser_sampled`

Class	Train	Val	Test	Total
wildebeestblue	1 244	267	266	1 777
zebraplains	919	197	197	1 313
gazellethomsons	724	155	156	1 035
buffalo	146	31	31	208
elephant	115	25	25	165
hartebeest	115	25	24	164
gazellegrants	74	16	16	106
warthog	66	14	15	95
impala	60	13	13	86
empty	35	8	7	50
Total	3 498	751	750	4 999

Statistics — `ser_sampled_cropped`

Same splits and class labels as ser_sampled. The empty class is excluded from the cropped variant (no animal detection box available). Total: 4 949 images.

Note: ~56% of images are IR/night (near-infrared, nearly greyscale).

Curation Details

Deduplication: one image per sequence (highest MD animal confidence frame)
Animal filter: MD animal confidence ≥ 0.8
Empty filter: max MD animal confidence < 0.2
Sampling: proportional to real Serengeti encounter rates (no per-class cap)
Split strategy: stratified 70/15/15 by sequence ID — no sequence spans splits
Image resolution: resized to max 1024 px on longer side, JPEG quality 92
Format: ImageFolder layout — <split>/<label>/<filename>.jpg

Usage

from huggingface_hub import hf_hub_download
import tarfile

# Full-frame version
archive = hf_hub_download(
    "marco-willi/ser_sampled",
    "ser_sampled.tar.gz",
    repo_type="dataset",
)
with tarfile.open(archive) as tar:
    tar.extractall(DATA_PATH)
# → DATA_PATH/ser_sampled/{train,val,test}/<label>/*.jpg

# MD-cropped version (recommended for Colab — smaller download)
archive = hf_hub_download(
    "marco-willi/ser_sampled_cropped",
    "ser_sampled_cropped.tar.gz",
    repo_type="dataset",
)
with tarfile.open(archive) as tar:
    tar.extractall(DATA_PATH)
# → DATA_PATH/ser_sampled_cropped/{train,val,test}/<label>/*.jpg