Aller au contenu
ModelPrésentéExterne
Browse files

Mamba Segmentation

Controlled Visual State-Space Backbone Benchmark with Domain-Shift & Boundary Analysis for Remote-Sensing Segmentation

Sourcehuggingface.co/dineth18/Mamba-Segmentation/tree/0df6efb255155c92a7fb5fd70731516fef4bd9dd
TypeOtherVersionRevision 0df6efbCClasse

Source du modèle

Type
Other
Taille du fichier
18,0 MB–2,78 GB
Version
vRevision 0df6efb
Mis à jour
16 sept. 2026
Téléchargements
0
Source
Hugging Face

Description de la source

Mamba-Segmentation

Controlled Visual State-Space Backbone Benchmark with Domain-Shift & Boundary Analysis for Remote-Sensing Segmentation

Accepted at IGARSS 2026

One pipeline. One decoder. One loss. One schedule. Five backbone families. The only variable is the encoder — so the results finally mean something.

What Is This?

Remote-sensing segmentation papers routinely change the backbone and the decoder and the loss and the training schedule all at once. The numbers tell you who tuned harder, not which backbone is better.

This repo fixes that. One shared pipeline — swap the backbone — read the truth.

ComponentStatus
Encoder backbone🔀 Swapped per experiment — the ONLY variable
Decoder🔒 Fixed (lightweight U-Net, 256ch, MambaBlock2d)
Loss🔒 Fixed (Lovász-Softmax + Focal + Boundary)
Training schedule🔒 Fixed (50k iters, AdamW, poly LR decay)
Augmentations🔒 Fixed (random crop, flip, colour jitter)
Input resolution🔒 Fixed (512×512)
Feature interface🔒 Fixed ({F1–F4} at strides {4, 8, 16, 32})

Checkpoints in This Repository

All checkpoints are best.pth files (highest validation mIoU during training) stored with their original directory structure.

LoveDA Experiments — Comparison_Experiments/

MambaVision (NVIDIA hybrid Mamba-Transformer)

Checkpoint pathTraining split
Comparison_Experiments/mambavision_tiny_512/checkpoints/best.pthAll→All
Comparison_Experiments/mambavision_tiny_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/mambavision_tiny_urbantrain_512/checkpoints/best.pthUrban→Rural
Comparison_Experiments/mambavision_tiny2_512/checkpoints/best.pthAll→All (v2)
Comparison_Experiments/mambavision_tiny2_ruraltrain_512/checkpoints/best.pthRural→Urban (v2)
Comparison_Experiments/mambavision_tiny2_urbantrain_512/checkpoints/best.pthUrban→Rural (v2)
Comparison_Experiments/mambavision_small_512/checkpoints/best.pthAll→All
Comparison_Experiments/mambavision_small_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/mambavision_small_urbantrain_512/checkpoints/best.pthUrban→Rural
Comparison_Experiments/mambavision_base_512/checkpoints/best.pthAll→All
Comparison_Experiments/mambavision_base_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/mambavision_base_urbantrain_512/checkpoints/best.pthUrban→Rural
Comparison_Experiments/mambavision_large_512/checkpoints/best.pthAll→All
Comparison_Experiments/mambavision_large_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/mambavision_large_urbantrain_512/checkpoints/best.pthUrban→Rural
Comparison_Experiments/mambavision_large2_512/checkpoints/best.pthAll→All
Comparison_Experiments/mambavision_large2_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/mambavision_large2_urbantrain_512/checkpoints/best.pthUrban→Rural

VMamba (cross-scan 2D selective SSM)

Checkpoint pathTraining split
Comparison_Experiments/Vmamb_tiny_512/checkpoints/best.pthAll→All
Comparison_Experiments/vmamba_tiny_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/vmamba_tiny_urbantrain_512/checkpoints/best.pthUrban→Rural
Comparison_Experiments/Vmamb_small_512/checkpoints/best.pthAll→All
Comparison_Experiments/Vmamb_small_512_2/checkpoints/best.pthAll→All (run 2)
Comparison_Experiments/Vmamb_small_512_3/checkpoints/best.pthAll→All (run 3)
Comparison_Experiments/vmamba_small_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/vmamba_small_urbantrain_512/checkpoints/best.pthUrban→Rural
Comparison_Experiments/Vmamb_base_512/checkpoints/best.pthAll→All
Comparison_Experiments/vmamba_base_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/vmamba_base_urbantrain_512/checkpoints/best.pthUrban→Rural

VisionMamba / Vim (bidirectional Mamba)

Checkpoint pathTraining split
Comparison_Experiments/VisionMamba_tiny_512/checkpoints/best.pthAll→All
Comparison_Experiments/visionmamba_tiny_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/visionmamba_tiny_urbantrain_512/checkpoints/best.pthUrban→Rural
Comparison_Experiments/VisionMamba_small_512/checkpoints/best.pthAll→All
Comparison_Experiments/visionmamba_small_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/visionmamba_small_urbantrain_512/checkpoints/best.pthUrban→Rural
Comparison_Experiments/VisionMamba_base_512/checkpoints/best.pthAll→All
Comparison_Experiments/visionmamba_base_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/visionmamba_base_urbantrain_512/checkpoints/best.pthUrban→Rural

Spatial-Mamba (spatially-aware SSM)

Checkpoint pathTraining split
Comparison_Experiments/spatialmamba_tiny_512/checkpoints/best.pthAll→All
Comparison_Experiments/spatialmamba_tiny_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/spatialmamba_tiny_urbantrain_512/checkpoints/best.pthUrban→Rural
Comparison_Experiments/spatialmamba_small_512/checkpoints/best.pthAll→All
Comparison_Experiments/spatialmamba_small_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/spatialmamba_small_urbantrain_512/checkpoints/best.pthUrban→Rural
Comparison_Experiments/spatialmamba_base_512/checkpoints/best.pthAll→All
Comparison_Experiments/spatialmamba_base_ruraltrain_512/checkpoints/best.pthRural→Urban
Comparison_Experiments/spatialmamba_base_urbantrain_512/checkpoints/best.pthUrban→Rural

CNN & Transformer Baselines

Checkpoint pathModel
Comparison_Experiments/cnn_deeplabv3p_r50_512/checkpoints/best.pthDeepLabv3+ ResNet-50, All→All
Comparison_Experiments/cnn_deeplabv3p_resnet50_ruraltrain_512/checkpoints/best.pthDeepLabv3+ ResNet-50, Rural→Urban
Comparison_Experiments/cnn_deeplabv3p_resnet50_urbantrain_512/checkpoints/best.pthDeepLabv3+ ResNet-50, Urban→Rural
Comparison_Experiments/cnn_unet_r50_512/checkpoints/best.pthU-Net ResNet-50, All→All
Comparison_Experiments/transformer_unetformer_r18_512/checkpoints/best.pthUNetFormer ResNet-18, All→All
Comparison_Experiments/transformerunetformer_resnet18_ruraltrain_512/checkpoints/best.pthUNetFormer ResNet-18, Rural→Urban
Comparison_Experiments/transformerunetformer_resnet18_urbantrain_512/checkpoints/best.pthUNetFormer ResNet-18, Urban→Rural

ISPRS Potsdam Experiments — Comparison_Experiments_ICPRS_potsdam/

Checkpoint pathModel
Comparison_Experiments_ICPRS_potsdam/mambavision_tiny_512/checkpoints/best.pthMambaVision-Tiny
Comparison_Experiments_ICPRS_potsdam/mambavision_tiny2_512/checkpoints/best.pthMambaVision-Tiny2
Comparison_Experiments_ICPRS_potsdam/mambavision_small_512/checkpoints/best.pthMambaVision-Small
Comparison_Experiments_ICPRS_potsdam/mambavision_base_512/checkpoints/best.pthMambaVision-Base
Comparison_Experiments_ICPRS_potsdam/mambavision_large_512/checkpoints/best.pthMambaVision-Large
Comparison_Experiments_ICPRS_potsdam/mambavision_large2_512/checkpoints/best.pthMambaVision-Large2
Comparison_Experiments_ICPRS_potsdam/vmamba_tiny_512/checkpoints/best.pthVMamba-Tiny
Comparison_Experiments_ICPRS_potsdam/vmamba_small_512/checkpoints/best.pthVMamba-Small
Comparison_Experiments_ICPRS_potsdam/vmamba_base_512/checkpoints/best.pthVMamba-Base
Comparison_Experiments_ICPRS_potsdam/spatialmamba_tiny_512/checkpoints/best.pthSpatial-Mamba-Tiny
Comparison_Experiments_ICPRS_potsdam/spatialmamba_small_512/checkpoints/best.pthSpatial-Mamba-Small
Comparison_Experiments_ICPRS_potsdam/spatialmamba_base_512/checkpoints/best.pthSpatial-Mamba-Base
Comparison_Experiments_ICPRS_potsdam/cnn_deeplabv3p_r50_512/checkpoints/best.pthDeepLabv3+ ResNet-50
Comparison_Experiments_ICPRS_potsdam/transformer_unetformer_r18_512/checkpoints/best.pthUNetFormer ResNet-18

ImageNet Backbone Weights — weights/imagenet/

FileDescription
weights/imagenet/resnet50-11ad3fa6.pthResNet-50 ImageNet-1K pretrained
weights/imagenet/resnet18-f37072fd.pthResNet-18 ImageNet-1K pretrained

Results Summary

Every row shares the same decoder, loss, optimizer, schedule, and data splits. The only variable is the encoder.

LoveDA

BackbonemIoU (All→All)mIoU (U→R)mIoU (R→U)
DeepLabv3+ ResNet-50 (CNN)43.0130.3639.98
UNetFormer ResNet-18 (Transformer)48.6134.5644.84
VMamba-Small 🥇55.6640.6253.52
MambaVision-Large55.2538.5354.01
Spatial-Mamba-Base48.0335.2346.55

ISPRS Potsdam

BackbonemIoU
DeepLabv3+ ResNet-5075.09
UNetFormer ResNet-1874.99
VMamba-Small 🥇77.59
MambaVision-Large77.07
Spatial-Mamba-Base70.00

Key findings:

  • SSMs outperform CNNs and Transformers by a significant margin under identical conditions (+7–12 mIoU on LoveDA).
  • Scaling the encoder past VMamba-Small yields diminishing returns under a fixed decoder.
  • Domain transfer is asymmetric across all backbone families (Rural→Urban consistently outperforms Urban→Rural by 10–15 points) — a data distribution property, not a model property.
  • Boundary accuracy collapses under domain shift while interior accuracy holds — every backbone, every family.

How to Load a Checkpoint

import torch

# Example: load MambaVision-Base best checkpoint for LoveDA All→All
ckpt = torch.load(
    "Comparison_Experiments/mambavision_base_512/checkpoints/best.pth",
    map_location="cpu"
)
# keys: 'model', 'optimizer', 'scheduler', 'iter', 'best_score'
model_state = ckpt["model"]

To build the full model and run inference, clone the code repository and follow the setup instructions:

git clone https://github.com/dineth18/Mamba-Segmentation
cd Mamba-Segmentation/MambaVision   # or VMamba/, spatial-mamba/, etc.
pip install -r requirements.txt

# Set your dataset path (no need to edit config files)
export LOVEDA_ROOT=/path/to/LoveDA
export POTSDAM_ROOT=/path/to/ISPRS_Potsdam

python eval.py --checkpoint path/to/best.pth

Citation

If this benchmark is useful for your research, please cite:

@article{wasalathilaka2026controlledbenchmark,
  title={A Controlled Benchmark of Visual State-Space Backbones with
         Domain-Shift and Boundary Analysis for Remote-Sensing Segmentation},
  author={Wasalathilaka, Nichula and Perea, Dineth and Samarakoon, Oshadha
          and Wijenayake, Buddhi and Godaliyadda, Roshan and Herath, Vijitha
          and Ekanayake, Parakrama},
  journal={IGARSS 2026},
  year={2026}
}

Acknowledgements

  • VMamba — Visual State Space Model
  • MambaVision — NVIDIA hybrid Mamba-Transformer
  • Spatial-Mamba — Spatially-aware Mamba
  • LoveDA — Land-cover domain adaptation dataset
  • [ISPRS...

Sources

1 source
Hugging FacePrincipalVersion 0df6efb255155c92a7fb5fd70731516fef4bd9dd
huggingface.co/dineth18/Mamba-Segmentation/tree/0df6efb255155c92a7fb5fd70731516fef4bd9dd

Vérifié 1 sept.

Download files

Artefacts du modèle

64 artefacts

Comparison_Experiments_ICPRS_potsdam/cnn_deeplabv3p_r50_512/checkpoints/best.pth

pth · 397 MB · SHA-256 27b5e13eaecd…c15e · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/mambavision_base_512/checkpoints/best.pth

pth · 1,16 GB · SHA-256 63d2abe09b62…4632 · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/mambavision_large_512/checkpoints/best.pth

pth · 2,63 GB · SHA-256 0b16b45e760c…003d · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/mambavision_large2_512/checkpoints/best.pth

pth · 2,78 GB · SHA-256 8696ef90463e…5321 · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/mambavision_small_512/checkpoints/best.pth

pth · 631 MB · SHA-256 9579b1dab4ef…d5f6 · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/mambavision_tiny_512/checkpoints/best.pth

pth · 417 MB · SHA-256 083678e42d2a…80c9 · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/mambavision_tiny2_512/checkpoints/best.pth

pth · 454 MB · SHA-256 cb46ea0dbd24…5f40 · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/spatialmamba_base_512/checkpoints/best.pth

pth · 1,12 GB · SHA-256 5c55c01fee44…098a · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/spatialmamba_small_512/checkpoints/best.pth

pth · 540 MB · SHA-256 9596d96dacad…79fe · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/spatialmamba_tiny_512/checkpoints/best.pth

pth · 352 MB · SHA-256 3acec83b50d8…c231 · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/transformer_unetformer_r18_512/checkpoints/best.pth

pth · 182 MB · SHA-256 b9ef0f4275bb…4738 · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/vmamba_base_512/checkpoints/best.pth

pth · 936 MB · SHA-256 a4986bd069e6…c513 · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/vmamba_small_512/checkpoints/best.pth

pth · 563 MB · SHA-256 fcc9b86bfef6…0e8e · Hugging Face

Download file

Comparison_Experiments_ICPRS_potsdam/vmamba_tiny_512/checkpoints/best.pth

pth · 316 MB · SHA-256 d97c4f13e8a8…1a4f · Hugging Face

Download file

Comparison_Experiments/cnn_deeplabv3p_r50_512/checkpoints/best.pth

pth · 397 MB · SHA-256 e3c0f80d7c44…5b87 · Hugging Face

Download file

Comparison_Experiments/cnn_deeplabv3p_resnet50_ruraltrain_512/checkpoints/best.pth

pth · 397 MB · SHA-256 4dd87a4ca634…e47f · Hugging Face

Download file

Comparison_Experiments/cnn_deeplabv3p_resnet50_urbantrain_512/checkpoints/best.pth

pth · 397 MB · SHA-256 db294a1e780d…8ad2 · Hugging Face

Download file

Comparison_Experiments/cnn_unet_r50_512/checkpoints/best.pth

pth · 397 MB · SHA-256 9bc6b9e7059e…32cc · Hugging Face

Download file

Comparison_Experiments/mambavision_base_512/checkpoints/best.pth

pth · 1,16 GB · SHA-256 8afaba1c4616…d4a5 · Hugging Face

Download file

Comparison_Experiments/mambavision_base_ruraltrain_512/checkpoints/best.pth

pth · 1,16 GB · SHA-256 a6b3508c83aa…45f9 · Hugging Face

Download file

Comparison_Experiments/mambavision_base_urbantrain_512/checkpoints/best.pth

pth · 1,16 GB · SHA-256 4945086da5fc…028b · Hugging Face

Download file

Comparison_Experiments/mambavision_large_512/checkpoints/best.pth

pth · 2,63 GB · SHA-256 b17ad656a5e6…bc61 · Hugging Face

Download file

Comparison_Experiments/mambavision_large_ruraltrain_512/checkpoints/best.pth

pth · 2,63 GB · SHA-256 4ca7a58f0835…2221 · Hugging Face

Download file

Comparison_Experiments/mambavision_large_urbantrain_512/checkpoints/best.pth

pth · 2,63 GB · SHA-256 6406bb454e16…3fe6 · Hugging Face

Download file

Comparison_Experiments/mambavision_large2_512/checkpoints/best.pth

pth · 2,78 GB · SHA-256 4f4cac9097f7…c318 · Hugging Face

Download file

Comparison_Experiments/mambavision_large2_ruraltrain_512/checkpoints/best.pth

pth · 2,78 GB · SHA-256 a3e287071317…4fdd · Hugging Face

Download file

Comparison_Experiments/mambavision_large2_urbantrain_512/checkpoints/best.pth

pth · 2,78 GB · SHA-256 e384fe38da26…352a · Hugging Face

Download file

Comparison_Experiments/mambavision_small_512/checkpoints/best.pth

pth · 631 MB · SHA-256 81901206b567…ec35 · Hugging Face

Download file

Comparison_Experiments/mambavision_small_ruraltrain_512/checkpoints/best.pth

pth · 631 MB · SHA-256 05a9c4d62426…de98 · Hugging Face

Download file

Comparison_Experiments/mambavision_small_urbantrain_512/checkpoints/best.pth

pth · 631 MB · SHA-256 a541b2d26a50…315e · Hugging Face

Download file

Comparison_Experiments/mambavision_tiny_512/checkpoints/best.pth

pth · 417 MB · SHA-256 b768d27cfb47…6cb1 · Hugging Face

Download file

Comparison_Experiments/mambavision_tiny_ruraltrain_512/checkpoints/best.pth

pth · 417 MB · SHA-256 16d3ed401125…de96 · Hugging Face

Download file

Comparison_Experiments/mambavision_tiny_urbantrain_512/checkpoints/best.pth

pth · 417 MB · SHA-256 b9f04d229d7d…a3eb · Hugging Face

Download file

Comparison_Experiments/mambavision_tiny2_512/checkpoints/best.pth

pth · 454 MB · SHA-256 e93d18fbbb4e…a4a8 · Hugging Face

Download file

Comparison_Experiments/mambavision_tiny2_ruraltrain_512/checkpoints/best.pth

pth · 454 MB · SHA-256 16b469cdf91f…9fb8 · Hugging Face

Download file

Comparison_Experiments/mambavision_tiny2_urbantrain_512/checkpoints/best.pth

pth · 454 MB · SHA-256 6d8dd7fbbcc6…21b2 · Hugging Face

Download file

Comparison_Experiments/spatialmamba_base_512/checkpoints/best.pth

pth · 1,12 GB · SHA-256 6a533634ad4e…61f5 · Hugging Face

Download file

Comparison_Experiments/spatialmamba_base_ruraltrain_512/checkpoints/best.pth

pth · 1,12 GB · SHA-256 dd8b3b38b355…90cb · Hugging Face

Download file

Comparison_Experiments/spatialmamba_base_urbantrain_512/checkpoints/best.pth

pth · 1,12 GB · SHA-256 5d1d74934c95…a649 · Hugging Face

Download file

Comparison_Experiments/spatialmamba_small_512/checkpoints/best.pth

pth · 540 MB · SHA-256 b17d3e2d6050…bc6a · Hugging Face

Download file

Comparison_Experiments/spatialmamba_small_ruraltrain_512/checkpoints/best.pth

pth · 540 MB · SHA-256 aadb4145afbb…02ae · Hugging Face

Download file

Comparison_Experiments/spatialmamba_small_urbantrain_512/checkpoints/best.pth

pth · 540 MB · SHA-256 02f38b54529c…045c · Hugging Face

Download file

Comparison_Experiments/spatialmamba_tiny_512/checkpoints/best.pth

pth · 352 MB · SHA-256 41e4ecc3fce5…3acd · Hugging Face

Download file

Comparison_Experiments/spatialmamba_tiny_ruraltrain_512/checkpoints/best.pth

pth · 352 MB · SHA-256 5ffb243fdafc…adee · Hugging Face

Download file

Comparison_Experiments/spatialmamba_tiny_urbantrain_512/checkpoints/best.pth

pth · 352 MB · SHA-256 89cf61149972…90df · Hugging Face

Download file

Comparison_Experiments/transformer_unetformer_r18_512/checkpoints/best.pth

pth · 182 MB · SHA-256 f4dc4500a1f5…7002 · Hugging Face

Download file

Comparison_Experiments/transformerunetformer_resnet18_ruraltrain_512/checkpoints/best.pth

pth · 182 MB · SHA-256 12a0feb9abeb…31ea · Hugging Face

Download file

Comparison_Experiments/transformerunetformer_resnet18_urbantrain_512/checkpoints/best.pth

pth · 182 MB · SHA-256 961700194baf…40d7 · Hugging Face

Download file

Comparison_Experiments/VisionMamba_base_512/checkpoints/best.pth

pth · 84,3 MB · SHA-256 2e68fb0f2a4c…1824 · Hugging Face

Download file

Comparison_Experiments/visionmamba_base_ruraltrain_512/checkpoints/best.pth

pth · 84,3 MB · SHA-256 8d798304242c…2e4f · Hugging Face

Download file

Comparison_Experiments/visionmamba_base_urbantrain_512/checkpoints/best.pth

pth · 84,3 MB · SHA-256 631a276836d1…dbd6 · Hugging Face

Download file

Comparison_Experiments/VisionMamba_small_512/checkpoints/best.pth

pth · 34,6 MB · SHA-256 a23b81730117…0f7f · Hugging Face

Download file

Comparison_Experiments/visionmamba_small_ruraltrain_512/checkpoints/best.pth

pth · 34,6 MB · SHA-256 a6272b95c2ed…97da · Hugging Face

Download file

Comparison_Experiments/visionmamba_small_urbantrain_512/checkpoints/best.pth

pth · 34,6 MB · SHA-256 f81160a208a1…2e5f · Hugging Face

Download file

Comparison_Experiments/VisionMamba_tiny_512/checkpoints/best.pth

pth · 18,0 MB · SHA-256 949af33b29be…c288 · Hugging Face

Download file

Comparison_Experiments/visionmamba_tiny_ruraltrain_512/checkpoints/best.pth

pth · 18,0 MB · SHA-256 fffba6e45e15…c7b1 · Hugging Face

Download file

Comparison_Experiments/visionmamba_tiny_urbantrain_512/checkpoints/best.pth

pth · 18,0 MB · SHA-256 6d8c9be7e8c6…7b63 · Hugging Face

Download file

Comparison_Experiments/Vmamb_base_512/checkpoints/best.pth

pth · 936 MB · SHA-256 ae190e957078…9f9b · Hugging Face

Download file

Comparison_Experiments/Vmamb_small_512_2/checkpoints/best.pth

pth · 563 MB · SHA-256 8ddf688e2567…7d70 · Hugging Face

Download file

Comparison_Experiments/Vmamb_small_512_3/checkpoints/best.pth

pth · 563 MB · SHA-256 e986d9269fb2…61e0 · Hugging Face

Download file

Comparison_Experiments/Vmamb_small_512/checkpoints/best.pth

pth · 563 MB · SHA-256 72d0f8e70b04…5d29 · Hugging Face

Download file

Comparison_Experiments/Vmamb_tiny_512/checkpoints/best.pth

pth · 316 MB · SHA-256 f7de1810fd7c…335e · Hugging Face

Download file

Comparison_Experiments/vmamba_base_ruraltrain_512/checkpoints/best.pth

pth · 936 MB · SHA-256 fba1ae85a323…4733 · Hugging Face

Download file

Comparison_Experiments/vmamba_base_urbantrain_512/checkpoints/best.pth

pth · 936 MB · SHA-256 04c39e68fbb0…a094 · Hugging Face

Download file

Extraits de sources

2 extraits
Readme1 sept.huggingface.co/dineth18/Mamba-Segmentation

--- license: mit language: - en tags: - remote-sensing - semantic-segmentation - mamba - state-space-model - vmamba - mambavision - spatial-mamba - pytorch - benchmark - loveda - isprs-potsdam - domain-adaptation datasets: - LoveDA - ISPRS-Potsdam pipeline_tag: image-segmentation --- # Mamba-Segmentation **Controlled Visual State-Space Backbone Benchmark with Domain-Shift & Boundary Analysis for Remote-Sensing Segmentation** > *Accepted at IGARSS 2026* One pipeline. One decoder. One loss. One schedule. **Five backbone families.** The only variable is the encoder — so the results finally mean something. --- ## What Is This? Remote-sensing segmentation papers routinely change the backbone *and* the decoder *and* the loss *and* the training schedule all at once. The numbers tell you who tuned harder, not which backbone is better. This repo fixes that. **One shared pipeline — swap the backbone — read the truth.** | Component | Status | |---|---| | Encoder backbone | 🔀 **Swapped** per experiment — the ONLY variable | | Decoder | 🔒 Fixed (lightweight U-Net, 256ch, MambaBlock2d) | | Loss | 🔒 Fixed (Lovász-Softmax + Focal + Boundary) | | Training schedule | 🔒 Fixed (50k iters, AdamW, poly LR decay) | | Augmentations | 🔒 Fixed (random crop, flip, colour jitter) | | Input resolution | 🔒 Fixed (512×512) | | Feature interface | 🔒 Fixed ({F1–F4} at strides {4, 8, 16, 32}) | --- ## Checkpoints in This Repository All checkpoints are `best.pth` files (highest validation mIoU during training) stored with their original directory structure. ### LoveDA Experiments — `Comparison_Experiments/` #### MambaVision (NVIDIA hybrid Mamba-Transformer) | Checkpoint path | Training split | |---|---| | `Comparison_Experiments/mambavision_tiny_512/checkpoints/best.pth` | All→All | | `Comparison_Experiments/mambavision_tiny_ruraltrain_512/checkpoints/best.pth` | Rural→Urban | | `Comparison_Experiments/mambavision_tiny_urbantrain_512/checkpoints/best.pth` | Urban→Rural | | `Comparison_Experiments/mambavision_tiny2_512/checkpoints/best.pth` | All→All (v2) | | `Comparison_Experiments/mambavision_tiny2_ruraltrain_512/checkpoints/best.pth` | Rural→Urban (v2) | | `Comparison_Experiments/mambavision_tiny2_urbantrain_512/checkpoints/best.pth` | Urban→Rural (v2) | | `Comparison_Experiments/mambavision_small_512/checkpoints/best.pth` | All→All | | `Comparison_Exp...

Note1 sept.huggingface.co/dineth18/Mamba-Segmentation

Source context: 0 downloads · 1 likes · Pipeline image-segmentation · Repo dineth18/Mamba-Segmentation