uav-segformer-b2-finetune-in-lowalt

UAV SegFormer-B2 fine-tuned for low-altitude imagery

This repository provides a ZIP archive containing the two verified checkpoints from the uavseg_segformer_b2_v2 experiment: a SegFormer-B2 model fine-tuned for 9-class low-altitude aerial semantic segmentation.

Files

• uavseg_segformer_b2_v2_checkpoints.zip — download this archive and extract it. It contains best.pt, last.pt, config.yaml, README.md, and checkpoint_manifest.json at its root.

Archive SHA-256: 3b656099c7d0b5940b313a99ba17527811920425268a3fef0c8eac78f253d1e0

Inside the archive, best.pt is selected by validation mIoU; last.pt is the complete epoch-40 checkpoint, including optimizer and scheduler state for resuming training. checkpoint_manifest.json contains the individual checkpoint hashes and provenance.

The checkpoint format is the project-specific uavseg PyTorch format. Load it with uavseg.utils.checkpoint.load_checkpoint, rather than directly with a generic Transformers from_pretrained call.

Inference

Use the reproduction code repository:

git clone https://github.com/wsffdg-123/uav-segformer-b2-finetune-in-lowalt.git
cd uav-segformer-b2-finetune-in-lowalt
# Install CUDA-compatible torch, then:
python -m pip install -r requirements/train.txt

After extracting the archive, copy best.pt into outputs/experiments/uavseg_segformer_b2_v2/best.pt, then run:

python scripts/predict.py \
  --config configs/optimized_v2.yaml \
  --checkpoint outputs/experiments/uavseg_segformer_b2_v2/best.pt \
  --image-dir /path/to/images \
  --output-dir outputs/predictions

The reproduction package deliberately provides only single-scale whole-image inference; it does not include TTA, sliding windows, or other inference post-processing. No dataset is distributed in this repository.

Provenance

• Source experiment: uavseg_segformer_b2_v2
• Source code revision: 965610272584e3f285fc88531c48cfe9d2812764
• Backbone initializer: nvidia/mit-b2 revision 3a609931044a9a83814802af9d861a47a1397636
• License: Apache-2.0

Please verify every downloaded checkpoint against checkpoint_manifest.json before use.