Skip the 3D Convolutions, Keep the Geometry

What it does

SimpleRecon is the reference PyTorch implementation for an ECCV 2022 multi-view stereo depth estimation method. Feed it posed RGB images and it outputs depth maps for target frames, which you can then fuse into meshes. It is essentially a batteries-included research codebase: pretrained weights, evaluation scripts, and dataset loaders for ScanNetv2, COLMAP, and ARKit scans are all included.

The interesting bit

The project’s name is its thesis: it reconstructs 3D scene geometry without ever touching a 3D convolution. Instead, it relies on 2D feature matching and cost volumes, a choice that keeps GPU memory around 2.6 GB for the standard model and inference latency as low as ~70 ms in speed-optimized mode.

Key highlights

• Reference implementation of the ECCV 2022 paper with two pretrained checkpoints (hero_model and dot_product_model).
• Lean inference footprint: the standard model needs roughly 2.6 GB of VRAM and runs at ~130 ms per frame, or ~70 ms in speed-optimized mode.
• Supports multiple data pipelines beyond ScanNetv2, including COLMAP reconstructions and ARKit/ios-logger captures.
• Ships with sample scans, depth-fusion tools, and mesh-evaluation scripts so you can go from images to a mesh without writing glue code.
• Non-commercial license (Niantic).

Caveats

• Strictly non-commercial license; commercial use is off the table.
• FPN model weights are still listed as a pending TODO.
• A late-2022 bug fix for intrinsics and projection precision improved scores but broke compatibility with old forks and weights.

Verdict

Grab it if you are an academic or hobbyist hacking on indoor MVS reconstruction and need a working baseline. Skip it if you need a production-ready or commercially licensable 3D pipeline.