---
pretty_name: Dynamic Multi-View Matching Benchmark
tags:
- synthetic
- blender
- multi-view
- feature-matching
- correspondence
- depth
- instance-segmentation
task_categories:
- image-feature-extraction
- keypoint-detection
- depth-estimation
- image-segmentation
license: other
language:
- en
---

# Evaluating Feature Matching in Dynamic Scenes: A Novel Dataset for Benchmarking

## Summary
This repository contains:
1) **A synthetic multi-view dataset** generated in Blender for benchmarking feature matching under **camera motion + scene dynamics** (object motion / addition-removal / scale changes / lighting changes).
2) **Evaluation outputs** obtained by running matching models on the dataset using [Glue Factory](https://github.com/cvg/glue-factory) framework.

---

## What is inside this repo

### A) Dataset (blender_dataset folder)
The dataset consists of multi-view scenes. Each scene provides:
- RGB renders (`.png`)
- Depth maps (`.exr`)
- Object masks (`.exr`)
- Calibrated camera parameters per view (`.npz`: intrinsics + extrinsics)
- Per-object pose/scale per view (`.npz`)

### B) Evaluation artifacts (outputs.zip file)
This repo also includes artifacts produced by running evaluation on:
- SuperGlue
- LoFTR
- LightGlue
- GlueStick