README.md

---
license: cc-by-nc-4.0
language:
- en
tags:
- image-classification
- bird-classification
- birds
- dinov2
- fine-grained
- north-america
library_name: transformers
pipeline_tag: image-classification
base_model: facebook/dinov2-base
datasets:
- yashikota/birds-525-species-image-classification
metrics:
- accuracy
---

# birdclass-na

A bird species classifier optimized for **North American backyard and
camera-trap conditions** — partial occlusion, motion blur, fence/leaf
clutter, low-light, and other things you don't see in handheld photo
datasets but you do see when a feeder camera is the photographer.

Backbone: [`facebook/dinov2-base`](https://huggingface.co/facebook/dinov2-base)
(Apache-2.0), with a `Linear(768, 407)` classification
head trained on a unified taxonomy spanning gpiosenka 525,
NABirds, iNat21-Birds, and the [BirdWatcher](https://github.com/houlette/BirdWatcher)
yard dataset.

## What this model is good at

- North American backyard bird identification, **especially** under
  feeder-camera conditions (partial occlusion, motion blur, leaf
  clutter, low-light) — the training mix includes ~2,000 user-labeled
  crops from a real feeder-camera deployment alongside the public
  datasets.
- Fine-grained discrimination of common NA confusables (Mourning Dove
  vs Rock Pigeon, Cooper's Hawk vs Sharp-shinned Hawk).

## What this model is _not_ good at

- **Non-NA species**: most non-NA bird images in iNat21 were collapsed
  into a single `OTHER` bucket during training. The model can flag a
  bird as "not one of these 406 NA species" but can't tell
  you _which_ non-NA species it is.
- **Rare-species long tail**: NA species with very few training samples
  (< 30 each) have low individual accuracy. We're not better than
  general-purpose bird classifiers there, just smaller.
- **Comparison to Cornell Merlin / iNat CV**: those are trained on
  10-100× more data and remain stronger in absolute terms on most
  common-species photos. This model's value is in being open-source,
  fine-tunable, and stronger on camera-trap conditions.

## Benchmarks

# BENCHMARK.md

Test set: **27,470 rows** held out from gpiosenka 525, NABirds, and the
BirdWatcher yard dataset. (No iNat21 test split — iNat21 only contributed
to train/val.)

All three models scored apples-to-apples in our **407-way canonical
taxonomy** (406 NA species + OTHER). Comparator outputs are mapped through
the same alias table — `Rock Dove` → `Rock Pigeon`,
`Cardinalis cardinalis` → `Northern Cardinal`, etc. — that our taxonomy
builder uses. denisjooo's 525-way logits and birder's 10,000-way logits
are max-pooled per canonical bucket; ours predict natively.

## Three-way: ours vs denisjooo vs birder-project

| Split | n | **Ours** | [denisjooo](https://huggingface.co/dennisjooo/Birds-Classifier-EfficientNetB2) | [birder-project](https://huggingface.co/birder-project/hieradet_d_small_dino-v2-inat21) |
|---|---:|---:|---:|---:|
| **overall** | 27,470 | **92.9%** (92.6–93.2) | 23.8% (23.3–24.4) | 89.6% (89.3–90.0) |
| gpiosenka | 2,625 | 89.0% (87.9–90.1) | 99.0% (98.7–99.4) | 85.3% (84.0–86.8) |
| nabirds | 24,633 | **93.3%** (93.0–93.6) | 15.9% (15.5–16.4) | 90.4% (90.0–90.7) |
| **yard** | 212 | **96.2%** (93.9–98.6) | 10.4% (6.6–14.6) | 57.1% (50.9–63.7) |

_Top-1 with 95 % bootstrap CIs over 1,000 resamples. Bold marks the best in
each row._

## How to read this

- **Overall**: we beat both alternatives. The +3.3 pp lead over birder
  is outside the CI overlap.
- **NABirds** (the cleanest NA-species test split, n=24,633): we beat
  birder by +2.9 pp on the source they'd be most expected to win.
- **Yard** (real feeder-camera crops with motion blur / partial
  occlusion / fence clutter, n=212): we beat birder by **+39.1 pp**.
  This is the validation of our "domain fine-tune on production yard
  data" thesis. Birder's iNat21-only training has no exposure to
  camera-trap conditions.
- **gpiosenka**: denisjooo wins (+10 pp over us) because gpiosenka's
  test split *is* its training data's holdout. We beat birder by +3.7 pp
  on this split despite the disadvantage.

## What this means for use cases

- **Best for backyard / feeder-camera / camera-trap conditions**: ours,
  by ~40 pp over the nearest competitor.
- **Best for clean handheld iNat-style photos**: birder is solid, especially
  if you also need plants / fungi / insects from the same model.
- **Best for the gpiosenka 525-species test specifically**: denisjooo
  (it was trained on those labels).
- **Best for "is this a bird I should care about?" with built-in NAB
  suppression**: ours (the `OTHER` class threshold gives a clean reject
  signal).


## Training data

- **gpiosenka 525**: ~89,885 images across 525 species. Pulled from
  [yashikota's HF mirror](https://huggingface.co/datasets/yashikota/birds-525-species-image-classification)
  (the original gpiosenka Kaggle upload was removed in 2025).
- **NABirds v1**: ~48,000 expert-labeled NA bird images from Cornell.
  Used under academic license — see https://dl.allaboutbirds.org/nabirds.
- **iNat21-Birds**: bird subset (~414k images) of the iNat 2021 challenge,
  filtered to the Aves supercategory. **License: CC-BY-NC**. This is
  why the trained model weights inherit a non-commercial restriction.
- **Yard data**: ~5,000 labeled crops from the [BirdWatcher](https://github.com/houlette/BirdWatcher)
  project. Domain-adaptation stage only.

## Quick start

```python
from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image

processor = AutoImageProcessor.from_pretrained("houlette/birdclass-na")
model = AutoModelForImageClassification.from_pretrained("houlette/birdclass-na")

img = Image.open("your_bird.jpg")
inputs = processor(images=img, return_tensors="pt")
outputs = model(**inputs)
top1 = outputs.logits.softmax(dim=-1)[0].argmax().item()
print(model.config.id2label[top1])
```

## Limitations and honest claims

This model is **best-in-class among open-source bird classifiers for
NA backyard / camera-trap use** — see the benchmark table above for
the head-to-head numbers vs the strongest alternatives we found
(denisjooo's EfficientNet and birder-project's Hiera-DINOv2-iNat21).
We win overall (+3.3 pp over birder, +69 pp over denisjooo) and by
~40 pp on real yard / camera-trap conditions.

It is **not** absolute SOTA on bird classification benchmarks. Cornell's
Merlin Bird ID app and iNaturalist's internal classifier are both
trained on orders of magnitude more data and remain stronger on most
common-species, clean-photo scenarios.

Use this model when:
- You need a local-running, fine-tunable bird classifier.
- Your inference distribution looks like camera-trap or feeder-camera
  imagery.
- You want Apache-2.0 code (the training pipeline) and CC-BY-NC
  weights with provenance you can audit.

Don't use this model when:
- You need commercial use (the iNat21 license restricts downstream).
  Re-train without iNat21 if commercial deployment matters.
- You need a global bird classifier — this is NA-focused by design.

## Citation

If you use this model in research, please cite it as:

```bibtex
@misc{birdclass_na_2026,
  author = {Houlette, Ryan},
  title = { birdclass-na: an open-source bird species classifier for North American backyards },
  year = { 2026 },
  publisher = { HuggingFace },
  url = { https://huggingface.co/houlette/birdclass-na }
}
```

## License

Apache-2.0 for the training pipeline at https://github.com/houlette/birdclass-na.
Model weights themselves are released under CC-BY-NC-4.0 due to inheritance
from iNat21's non-commercial clause.