You can deploy this model and start earning money today!

sulagnasaharasha/
bci-crown-model

Name: bci-crown-model
Availability: InStock

Views

No views yet

BCI Crown Classifier — DINOv3 ViT-B/16

Fine-tuned species classifier for tropical tree crowns from the Barro Colorado Island (BCI) dataset. The model predicts one of 84 tropical tree species from aerial crown-view RGB tiles (512×512 px).

Model Details


Backbone	DINOv3 ViT-B/16 (`dinov3_vitb16`), pretrained on LVD-1689M
Classifier	Linear head (768 → 84), dropout 0.1
Fine-tuning	Full model (backbone unfrozen), AdamW, lr=1e-4, wd=1e-4
Experiment	exp11, fold 0, epoch 14
Input size	512×512 RGB, ImageNet normalization

Performance (fold 0 test set)

Top-1	Top-3	Top-5
63.5%	76.0%	80.6%

DINOv3 Backbone

This model uses DINOv3 — a family of versatile vision foundation models trained with a self-supervised method that produces high-quality dense features across a broad range of vision tasks.

The backbone used here is dinov3_vitb16, pretrained on the LVD-1689M web dataset.
The backbone code is included in this repo under dinov3_repo/ and is loaded via torch.hub.
See dinov3_repo/MODEL_CARD.md for full DINOv3 model details.

Note: The dinov3_repo/ folder in this HuggingFace repo contains the DINOv3 source code needed to instantiate the backbone via torch.hub.load(..., source="local"). This is a local copy required because DINOv3 is not yet available on the standard torch.hub registry.

Repository Contents

bci-crown-model/
├── dinov3-14-0.000.ckpt              # Fine-tuned Lightning checkpoint (backbone + classifier)
├── dinov3_vitb16_pretrain_lvd1689m   # DINOv3 ViT-B/16 pretrained backbone weights
├── random_split_species_84.json      # List of 84 class names (index = model output index)
└── dinov3_repo/                      # DINOv3 source code for torch.hub.load
    ├── hubconf.py
    ├── dinov3/
    └── ...

Usage

python

from pathlib import Path
from huggingface_hub import hf_hub_download, snapshot_download
import torch
from bci_crown_classifier.models.dinov3_classifier import DinoV3LinearClassifier

HF_REPO = "sulagnasaharasha/bci-crown-model"

ckpt_path    = hf_hub_download(HF_REPO, "dinov3-14-0.000.ckpt")
weights_path = hf_hub_download(HF_REPO, "dinov3_vitb16_pretrain_lvd1689m")
repo_dir     = snapshot_download(HF_REPO, allow_patterns="dinov3_repo/**")
repo_path    = Path(repo_dir) / "dinov3_repo"

model = DinoV3LinearClassifier.load_from_checkpoint(
    ckpt_path,
    map_location="cpu",
    strict=False,
    weights_path=weights_path,
    repo_path=str(repo_path),
)
model.eval()

See the tutorial notebook in the bci-crown-classifier repository for a full end-to-end inference example.

Classes

84 tropical tree species from BCI, Panama. Full list in random_split_species_84.json.

Citation

If you use this model, please cite the BCI crown classifier work and the DINOv3 paper:

bibtex

@misc{dinov3,
  title  = {DINOv3: Learning Robust Visual Features without Supervision},
  author = {Oquab, Maxime and others},
  year   = {2024},
}