SPECTRA checkpoints
Image matchers fine-tuned on Spectra dataset build with RASMD, a cross-spectral RGB-SWIR driving dataset. Each folder holds one checkpoint. Architectures are unchanged from upstream; only the weights differ, so each loads with its original codebase.
| Folder | Model | Base | File | Size |
|---|---|---|---|---|
loma/ |
LoMa fine-tuned on RASMD RGB-SWIR | LoMa-B | avg_weights.pth |
0.71 GB |
mast3r/ |
MASt3R fine-tuned on RASMD RGB-SWIR | MASt3R ViT-L | checkpoint-best.pth |
2.57 GB |
mast3r_fast/ |
MASt3R fine-tuned on RASMD, fp16 fast-inference weights | MASt3R ViT-L | checkpoint-best_fast.pth |
1.30 GB |
minima_roma/ |
MINIMA-RoMa fine-tuned on RASMD with MegaDepth replay | MINIMA-RoMa | train_roma_rsmd_latest.pth |
1.24 GB |
roma/ |
RoMa fine-tuned on RASMD RGB-SWIR | RoMa (roma_outdoor) | train_roma_rsmd_latest.pth |
1.24 GB |
roma_mix/ |
RoMa fine-tuned on RASMD with MegaDepth replay | RoMa (roma_outdoor) | train_roma_rsmd_latest.pth |
1.24 GB |
Usage
from huggingface_hub import hf_hub_download
path = hf_hub_download("shahzeen/spectra_checkpoints", "roma/train_roma_rsmd_latest.pth")
mast3r_fast holds the same weights as mast3r stored in fp16, for a reduced-latency
inference path; it is not a smaller model.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support