Instructions to use AIhdproduction/Flurseg with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use AIhdproduction/Flurseg with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-segmentation", model="AIhdproduction/Flurseg")# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("AIhdproduction/Flurseg", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Flurseg
Built with DINOv3.
Flurseg is a top-down semantic segmentation model for drone, orthophoto and phone imagery. It assigns every pixel to one of 11 classes for site and urban surface mapping (pavement, ground, vegetation, buildings, vehicles, construction objects, curbs, road markings, utility covers).
The model is a student distilled from SAM 3: SAM 3 automatically generated the initial pseudo-labels, which were then manually reviewed, corrected and refined to improve segmentation accuracy and label quality. An EoMT-DINOv3-Large model was subsequently fine-tuned on these human-refined labels. The student needs only one forward pass per tile instead of multiple text-prompted SAM 3 passes.
Each model version lives in its own subfolder.
Versions
| Version | Folder | Date | Val mIoU* | Notes |
|---|---|---|---|---|
| v1 (recommended) | v1/ |
2026-11-08 | 0.6459 | Trained on SAM 3 pseudo-labels, manually reviewed and refined; 40,000 steps |
| v2 | v2/ |
2026-11-10 | 0.9792 (see note) | Self-training round 1, experimental: not better than v1 on the held-out test, see evaluation/ |
*mIoU on the validation split, measured against the manually refined SAM 3 labels of that split (not an independent, separately annotated test set).
Note: the v2 value is measured against round-1 masks fused from v1's own predictions and SAM 3, so it is not comparable to v1's. On an independent test of 9 new images v2 is not better than v1 (see Evaluation below).
Evaluation
Independent comparison on 9 new, freely licensed images (3 drone, 3 orthophoto, 3 phone) that were not in the training data. The ground truth is the SAM 3 pre-label, manually corrected by the author; metrics on all pixels with GT not equal to 255.
| Model | Pixel acc | mIoU | Runtime per image |
|---|---|---|---|
| SAM 3 | 0.960 | 0.833 | 45.6 s |
| v1 | 0.913 | 0.732 | 6.0 s |
| v2 | 0.907 | 0.717 | 6.0 s |
The students are about 7.6 times faster than SAM 3; v2 is not better than v1. Caveat: the ground truth is derived from corrected SAM 3 masks and therefore favours SAM 3, and 9 images are too few to separate v1 from v2. Per-class results, comparison panels and the protocol: evaluation/README.md.
How SAM 3 was used
SAM 3 served as label teacher and as the reference in the evaluation. Weights, software versions, per-class text prompts, fusion priority, pipeline steps and the full configuration (local profile and the GPUHub profile used for the training labels) are documented in evaluation/SAM3_SETUP.md. The SAM 3 weights are not distributed here.
Quick start
import numpy as np, torch
from transformers import EomtDinov3ForUniversalSegmentation
model = EomtDinov3ForUniversalSegmentation.from_pretrained(
"AIhdproduction/Flurseg", subfolder="v1"
).eval().cuda()
Preprocessing: RGB, scaled to 0-1, normalised with ImageNet mean [0.485, 0.456, 0.406] and std [0.229, 0.224, 0.225].
Do not downscale large images. Use a sliding window: window 768 px, stride 384 px, accumulate the per-class scores of the overlapping windows and take the per-pixel argmax over the 11 classes. See v1/README.md for details.
Classes
| ID | Class | Colour |
|---|---|---|
| 0 | other | #9E9E9E |
| 1 | ground | #A0522D |
| 2 | paved_surface | #5C6BC0 |
| 3 | low_vegetation | #9CCC65 |
| 4 | tree | #1B5E20 |
| 5 | building | #E53935 |
| 6 | vehicle | #FFEB3B |
| 7 | construction | #FF9800 |
| 8 | curb | #00BCD4 |
| 9 | road_marking | #FFFFFF |
| 10 | utility_cover | #E040FB |
Label value 255 means "ignore" (used only in training labels, never predicted). Full definitions are in v1/classes.yaml.
Limitations
- The reported metric measures agreement with SAM 3 pseudo-labels, not directly with independently verified ground truth. However, all generated masks were manually reviewed and improved to correct segmentation errors and increase label quality. This reduces the impact of teacher errors and improves the reliability of the training data.
- Thin and rare classes (curb, road_marking, utility_cover) remain more challenging than frequent classes and may still perform worse than SAM 3, despite manual mask refinement.
- Training imagery is mostly from Switzerland and other European countries (Norway, France, Germany, Netherlands, Spain, Luxembourg, Sweden), with a small number of orthophotos from New Zealand, the USA and other regions. Performance may vary across different regions and sensors.
- Intended for top-down views (drone, orthophoto, phone looking down).
License
The model weights are released under the DINOv3 License (see LICENSE), because the backbone is DINOv3. Built with DINOv3. The base checkpoint tue-mps/eomt-dinov3-ade-semantic-large-512 (EoMT, MIT) comes with an ADE20K-pretrained head. SAM 3 was used only as a label teacher; SAM 3 weights are not distributed here. The training data is not distributed. See NOTICE.md for third-party notices and data attribution.
Acknowledgements and citation
- DINOv3: Simeoni et al., "DINOv3", Meta AI, 2025.
- EoMT: Kerssies et al., "Your ViT is Secretly an Image Segmentation Model", CVPR 2025.
- SAM 3: Meta AI, "SAM 3: Segment Anything with Concepts", 2025.
Model tree for AIhdproduction/Flurseg
Base model
tue-mps/eomt-dinov3-ade-semantic-large-512