ORBIT-Mamba / MambaCD checkpoints

Validation-selected MambaCD checkpoints for binary, bi-temporal remote-sensing change detection. Each release contains the checkpoint, the exact saved training configuration, the original test metrics, and a normalized metrics.json.

All accuracy metrics are fractions in [0, 1]. The F1 score is for the changed class. mIoU is the mean of changed and unchanged class IoU. GPU memory is the peak reserved training memory recorded in the source run, not inference-only memory.

Complete output and ablation archive

The repository also contains one validation-selected best checkpoint from every discovered training run: 142 models / 88.40 GiB before Xet deduplication. This includes all 45 standard outputs runs and all 97 ablation runs across the general, WHU, progressive LEVIR, and DRBI feature-operation suites.

See RUNS.md for the complete metrics table and direct model download links, or run-index.json for the machine-readable index. Every runs/.../ directory includes model.pth, normalized metrics.json, and the original train.log; source configs and raw metric histories are included wherever the run emitted them.

Released models

Dataset	Backbone	F1	mIoU	Overall accuracy	Recall	Precision	BF1	BmIoU	GFLOPs	GPU GB	FPS	Parameters	Download
LEVIR-CD+	VMamba-base	0.8728	0.8817	0.9895	0.8815	0.8642	0.8321	—	—	12.23	—	96.0M	model.pth
WHU-CD	VMamba-base	0.9617	0.9612	0.9963	0.9573	0.9662	0.9321	—	—	12.23	—	96.0M	model.pth
WildFireS2	VMamba-tiny	0.8645	0.6477	0.8126	0.9080	0.8250	—	—	—	8.68	—	44.0M	model.pth
SYSU-CD	VMamba-base	0.8457	0.8213	0.9279	0.8382	0.8533	0.4600	—	—	12.23	—	96.0M	model.pth
DSIFN-CD	VMamba-small	0.7127	0.7223	0.9039	0.7012	0.7245	0.3705	—	—	10.47	—	57.2M	model.pth
KATE-CD-256	VMamba-small	0.6612	0.7376	0.9816	0.6357	0.6889	0.6411	—	—	10.61	—	57.9M	model.pth

— means the value was not measured in the source run. BmIoU was not implemented by the evaluator. GFLOPs and FPS were requested in the saved configs but were not emitted by the runs, and this release does not estimate them. The WildFireS2 run predates the corrected symmetric boundary evaluator, so its recorded BF1 value of 0.0 is retained in source_test_metrics.json but treated as unavailable in the normalized table.

Download

Download one checkpoint with Python:

from huggingface_hub import hf_hub_download

checkpoint = hf_hub_download(
    repo_id="dineth18/ORBIT-Mamba",
    filename="models/whu-cd/model.pth",
)

Or clone every checkpoint (git-xet is recommended for the large files):

git xet install
git clone git@hf.co:dineth18/ORBIT-Mamba

Repository layout

models/<dataset>/
  model.pth                # validation-selected checkpoint
  config.yaml              # exact configuration saved by the run
  resolved_config.yaml     # included when the run emitted one
  source_test_metrics.json # unmodified evaluator output
  metrics.json             # normalized release metrics and provenance
  train.log                # original training log

See model-index.json for a machine-readable index and direct download URLs.

Loading

Use the matching MambaCD source code and the included configuration. Checkpoints are PyTorch training checkpoints rather than standalone TorchScript or ONNX exports. Local dataset, cache, and pretrained-weight paths in the saved configs document the original run and must be changed for a new machine.

import torch

state = torch.load("models/whu-cd/model.pth", map_location="cpu", weights_only=False)
print(state.keys())

Metric provenance

• F1, mIoU, overall accuracy, recall, precision, and BF1 come directly from each run's metrics/test_metrics.json.
• Thresholds were selected on validation data and then applied to the test split.
• Parameter totals come from the model summary written during training.
• GPU GB is the maximum gpu_memory value in the run's metrics/train_history.json and represents peak reserved training memory.
• BmIoU, GFLOPs, and FPS are explicitly null when no reproducible measurement exists.

License

This Hugging Face repository is marked MIT. The source MambaCD repository did not contain a root license file in the inspected checkout, so users must also review the MambaCD, VMamba, pretrained-weight, and dataset terms before redistribution or commercial use.