DeepNet Potsdam segmentation weights
This Hub repository contains the final single-model state dictionaries for the four segmentation systems compared in DeepNet's Potsdam research portfolio. Files use safetensors; they contain model tensors only, without optimizer state, source imagery, labels, or predictions.
Models and observed benchmark results
| Model | Input recipe | OA | Six-class mIoU | Weight file | Terms |
|---|---|---|---|---|---|
| DINOv3 ViT-B/16 | Verified RGB from RGBIR | 92.0397% | 81.0932% | models/dinov3-vitb16-rgb/model.safetensors |
DINOv3 License Agreement |
| DINOv2 ViT-B/14 | Historical 3-band Potsdam input catalogued as IRRG | 91.4785% | 79.6122% | models/dinov2-vitb14-irrg/model.safetensors |
Apache-2.0 |
| PSPNet ResNet101 | Verified RGB | 91.2834% | 79.4235% | models/pspnet-resnet101-rgb/model.safetensors |
Apache-2.0 project weights; upstream notices apply |
| SAM ViT-B + LoRA | Historical 3-band Potsdam input catalogued as IRRG | 87.8142% | 73.7024% | models/sam-vitb-lora-irrg/model.safetensors |
Apache-2.0 adapter weights |
All metrics use the same historical 14-scene cohort and 467,862,855 scored pixels with six-class eroded references. The cohort was examined in prior campaigns; these are observed results, not a newly collected blind test or an SOTA claim. These scores compare complete recipes, and only the current RGB DINOv3/PSPNet runs share a matched cache and inference policy.
The DINOv2 and SAM inputs are the historical three-band files configured as IRRG; source audit found their stored values were NIR/green/blue. They should not be described as natural RGB. DINOv3 and PSPNet use verified RGB from the four-band RGBIR source.
Download and load
Install the safe-tensor and Hub clients:
pip install safetensors huggingface_hub
Download a state dictionary, then load it into the matching model implementation and model specification:
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
weights_path = hf_hub_download(
repo_id="damminhtien/deepnet-potsdam-segmentation",
filename="models/dinov3-vitb16-rgb/model.safetensors",
)
state_dict = load_file(weights_path, device="cpu")
# Construct the matching DeepNet model from its model spec, then:
# model.load_state_dict(state_dict, strict=True)
The model specs, study configs, and evaluation protocol for DINOv2/DINOv3 are included beside their weights. PSPNet and SAM construction details are in model-info.json and the linked source reports. The DeepNet repository contains the corresponding implementation and protocols. SAM's file contains only trainable adapter/decoder tensors; obtain the official SAM ViT-B base checkpoint separately and verify its SHA-256 against model-info.json.
Licenses and dataset terms
Licenses are model-specific because this bundle includes weights derived from several upstream models. The exact DINOv3 License Agreement is included in licenses/DINOv3-LICENSE.md and must accompany any redistribution of DINOv3-derived weights. DINOv2, SAM, and project PSPNet weights use Apache-2.0 as specified in their model cards. The TorchVision BSD notice is also included; TorchVision notes that pretrained weights may have separate source-data terms. Read every applicable license before use or redistribution.
The ISPRS source data are not included. Access imagery and labels directly from the official ISPRS portal and follow its current terms. Publishing these weights does not grant rights to the source dataset. See the included manifest.json for source-checkpoint hashes, exported file hashes, parameter tensor counts, and evidence links.