FDS-UNeXT

Checkpoints for "Boundary Fidelity in Frozen Foundation-Model Adaptation: When It Holds and When It Fails for Fine-Grained Rooftop Segmentation" (submitted to Remote Sensing).

FDS-UNeXT adapts two frozen foundation models, the SAM 2.1 Hiera-L encoder and DINOv2 ViT-L/14, to four-class rooftop material segmentation with 6.97 M trainable parameters (1.33 % of 523.1 M).

Code: https://github.com/Zhao-Yuancheng/fds-unext

Format: trainable tensors only

A full checkpoint is 2.0 GB, but 516.2 M of its 523.1 M parameters are the frozen trunks and are identical to the public SAM 2.1 and DINOv2 weights. Each file here stores only the 6.97 M tensors that training updated, which makes it about 28 MB. The round trip is exact: merging reproduces every tensor of the original file.

You still need the two public trunks:

wget -P weights https://dl.fbaipublicfiles.com/segment_anything_2/092824/sam2.1_hiera_large.pt
wget -P weights https://dl.fbaipublicfiles.com/dinov2/dinov2_vitl14/dinov2_vitl14_pretrain.pth

Files

File Trained on Used for
lrmd-fold0.pth … lrmd-fold4.pth LRMD, 0.14 m, five folds Table 3, Figures 2 and 5
potsdam-seed42.pth, potsdam-seed3407.pth ISPRS Potsdam, 0.05 m Table 4, Figure 9a
egcs-seed42.pth, egcs-seed3407.pth, egcs-seed666.pth E-GCS, 0.3 m Table 4, Figures 3 and 9b

Usage

hf download Jim-myoenk/FDS-UNeXT lrmd-fold0.pth --local-dir deltas

python tools/merge_trainable.py \
  --delta deltas/lrmd-fold0.pth \
  --hiera_path weights/sam2.1_hiera_large.pt \
  --dinov2_path weights/dinov2_vitl14_pretrain.pth \
  --out runs/fold0/FDS-UNeXT-best.pth

python eval.py \
  --hiera_path weights/sam2.1_hiera_large.pt \
  --dinov2_path weights/dinov2_vitl14_pretrain.pth \
  --checkpoint runs/fold0/FDS-UNeXT-best.pth \
  --split_path dataset/roof_dataset_v2/split_fold0.json --split test \
  --num_classes 4 --bf_tolerance 8 --dy_edge_offset 6 \
  --class_names background,high-albedo-steel,low-albedo-steel,tile

Classes: 0 background (concrete roofs, asphalt, soil, vegetation, roads, water, vehicles), 1 high-albedo color-coated steel, 2 low-albedo color-coated steel, 3 tile.

Results

Dataset Setting mIoU (%) Boundary F-score (%)
LRMD five folds, 0.14 m 70.00 ± 6.58 68.58 ± 5.08
Potsdam two seeds, 0.05 m 75.56 ± 0.49 73.18 ± 0.65
E-GCS three seeds, 0.3 m 95.15 ± 0.37 73.05 ± 1.30

The boundary F-score uses an 8-pixel tolerance and class-equal weighting (per-class mean over images, then mean over classes). Metrics skip classes that are absent from the reference tile.

For continuity with earlier releases, note that those releases used element-weighted BF (Potsdam seed 42: 73.14; E-GCS three-seed: 76.17 ± 0.59). The two conventions must not be mixed or subtracted.

Citation

@article{zhao2026fdsunext,
  title   = {Boundary Fidelity in Frozen Foundation-Model Adaptation: When It
             Holds and When It Fails for Fine-Grained Rooftop Segmentation},
  author  = {Zhao, Yuancheng and Wang, Jie and Li, Faguo and Liu, Erte and
             Meng, Xiangying and Chen, Yueyue and Yang, Jiapeng and
             Zhou, Yibo and Li, Qingxue},
  journal = {Remote Sensing},
  year    = {2026},
  note    = {Under review}
}

License: Apache-2.0 for this model and for the two upstream trunks.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support