loftr

Native Rust/tch LoFTR feature matching.

What This Crate Includes

• LoFTR config presets
• native model construction in Rust with tch
• loading Kornia-compatible state dict weights
• pairwise feature matching
• optional debug-stage dumps for validation

Public API

• LoftrConfig for preset and custom model configuration
• LoftrModel for model construction, weight loading, and inference
• LoftrMatches for match outputs
• LoftrDebugStages for validation-oriented stage summaries
• normalize_loftr_image for converting supported image layouts into LoFTR input tensors
• LoftrError for public error handling

Upstream References

• LoFTR upstream
• Kornia upstream
• kornia-rs upstream

Weights

This crate does not ship pretrained weights.

Generate them locally from the workspace root:

./scripts/generate_loftr_state_dict.sh

Then load them from:

artifacts/weights/loftr_outdoor_state_dict.safetensors

Example

use loftr::{LoftrConfig, LoftrModel};
use tch::{Device, Kind, Tensor};

let mut model = LoftrModel::new(Device::Cpu, LoftrConfig::outdoor())?;
model.load_weights("artifacts/weights/loftr_outdoor_state_dict.safetensors")?;

let image0 = Tensor::rand([1, 1, 128, 128], (Kind::Float, Device::Cpu));
let image1 = Tensor::rand([1, 1, 128, 128], (Kind::Float, Device::Cpu));
let matches = model.forward(&image0, &image1)?;

println!("match count = {}", matches.confidence.size()[0]);
# Ok::<(), loftr::LoftrError>(())

Docs.rs

The published API docs are built with the doc-only feature so docs.rs can render the crate without linking libtorch.