dinov3_onnx_classifier

DINOv3 geo pre-filter

A binary image classifier that runs before GeoCLIP geo-estimation. It decides whether an image shows a place that can plausibly be located, such as streets, buildings, landscape or landmarks. Images without one go no further, so GeoCLIP never assigns a location to them. That covers portraits and selfies, close-ups, screenshots, logos, documents and other graphics.

Files

Both files have the same interface, so either is a drop-in replacement for the other.

Interface

• Input pixel_values: float32 [batch, 3, 224, 224]. RGB, resized to 224×224 without keeping aspect ratio or cropping (bilinear), scaled to [0, 1], then normalised with ImageNet mean [0.485, 0.456, 0.406] and std [0.229, 0.224, 0.225].
• Output classification: float32 [batch, 2] logits. Index 0 means usable for geo-estimation, index 1 means reject.
• Decision: take the argmax. The operating threshold is built into the head's bias, so callers need no threshold of their own. Apply a softmax for probabilities.
• ONNX opset 17. Tested with onnxruntime 1.22.0.

import numpy as np
import onnxruntime as ort
from PIL import Image

session = ort.InferenceSession("25.09.25-dinov3_vitb16_geo_filter.onnx")
image = Image.open("photo.jpg").convert("RGB").resize((224, 224), Image.BILINEAR)
x = (np.asarray(image, np.float32) / 255 - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225]
logits = session.run(None, {"pixel_values": x.transpose(2, 0, 1)[None].astype(np.float32)})[0][0]
usable_for_geo = int(np.argmax(logits)) == 0

Model (25.09.25)

• Backbone: facebook/dinov3-vitb16-pretrain-lvd1689m, frozen.
• Features: the CLS token concatenated with the mean of the patch tokens (register tokens excluded), 1536 values.
• Head: dropout → linear 1536→512 → GELU → dropout → linear 512→2. Feature standardisation is folded into the first layer's weights.
• Trained with class-weighted cross-entropy and AdamW (lr 3e-4, weight decay 0.05, cosine schedule), early-stopped on validation average precision for the usable class.

Training data

• 40,000 de-duplicated images sampled at random from an internal collection of user-supplied images. The images are not published.
• Each image was labelled by SmolVLM2-2.2B-Instruct. It answered two yes/no questions, read from next-token probabilities:
  • whether the photo shows enough of a real place to guess where in the world it was taken;
  • whether a person, face, animal or object fills most of the picture.
• A logistic regression fitted on 200 hand-labelled images combined the two answers into a score with two cut-offs. Images between the cut-offs were dropped as uncertain.
• 36,163 labelled images remained, split by image hash:

Evaluation

VLM-labelled test split (5,379 images, 439 usable): ROC AUC 0.995, average precision 0.963. At the built-in threshold, 89.6% of the images it passes are usable, and it keeps 90.7% of the usable images.

Hand-labelled held-out set: 300 random images, never used for training or tuning. Ambiguous images are excluded from these counts.

Fresh unlabelled images (3,000, never seen in training): the previous filter passes 60.1% of them and the new one 11.9%.

Known weakness: people posed outdoors with some scenery behind them, such as group photos or a selfie on a street, are the most common remaining false passes. Dense street crowds can also pass.

Runtime

Per image, batch 1, onnxruntime 1.22.0 on CPU (AMD EPYC 7401P), median:

The new backbone needs about 4× the compute of the previous one. GPU latency was not measured.

License

Both files contain DINOv3 weights, so they are distributed under the DINOv3 License.