How RoMa v2 Achieves Harder, Better, Faster, Denser Feature Matching

Background

Dense feature matching seeks a pixel‑wise correspondence between two images captured from different viewpoints or under varying illumination. Prior approaches either achieve high accuracy at the expense of speed and memory (e.g., RoMa) or run fast but degrade on challenging scenes (e.g., UFM). RoMa v2 is built to combine high precision, fast inference, and low memory usage.

Method

Two‑stage matching‑then‑refinement pipeline

The system first produces a coarse dense correspondence and then refines it to sub‑pixel accuracy.

Coarse matcher

Backbone upgraded from frozen DINOv2 to DINOv3, which provides more robust representations under extreme illumination and modality changes.

The original Gaussian‑process regressor is replaced by a single‑head attention module that aggregates multi‑view context.

An auxiliary negative‑log‑likelihood (NLL) loss is added to encourage correct patch‑to‑patch matching.

Refinement stage

Custom CUDA kernel computes local correlation without allocating extra buffers, dramatically reducing GPU memory consumption.

The refiner operates at three scales (4×, 2×, 1×) with a UNet‑like architecture; channel dimensions are constrained to powers of two for computational efficiency.

Training strategy

Mixed‑dataset curriculum comprising ten datasets (MegaDepth, AerialMD, FlyingThings3D, etc.) to cover wide‑baseline, small‑baseline, dynamic‑object, and cross‑modal scenarios.

During training a systematic sub‑pixel bias is observed; the final model is obtained by keeping an exponential moving average (EMA) of the weights, which eliminates this bias.

The network also predicts a per‑pixel 2×2 covariance matrix, providing uncertainty estimates for downstream tasks.

Experiments

On 640×640 inputs RoMa v2 processes 30.9 image pairs / s—1.7× faster than the original RoMa—while using slightly less GPU memory. Compared with UFM it consumes far less memory with comparable throughput.

Relative pose estimation on MegaDepth‑1500 and ScanNet‑1500 shows higher AUC than all prior methods (RoMa, UFM, LightGlue). Dense‑matching metrics (average endpoint error and PCK) on six benchmarks (TA‑WB, MegaDepth, ScanNet++, etc.) are best‑in‑class. The model also generalises to cross‑modal challenges such as WxBS and the newly introduced SatAst benchmark, where it outperforms competitors by a large margin.

Code

The implementation and pretrained models are publicly released at https://github.com/Parskatt/romav2.