YoYo000/BlendedMVS

BlendedMVS: A Large-scale Dataset for Generalized Multi-view Stereo Networks

What it solves

BlendedMVS provides a large-scale dataset designed to train and evaluate generalized multi-view stereo (MVS) networks. It addresses the need for diverse, high-quality training data covering a wide range of scenes—such as architectures, sculptures, and small objects—to improve the robustness and generalization of 3D reconstruction algorithms.

How it works

The project provides several versions of the dataset (BlendedMVS, BlendedMVS+, and BlendedMVS++), which are structured to be compatible with the MVSNet input format. The data consists of blended images (with and without masks), camera parameters (cams), and rendered depth maps. A superset called BlendedMVG expands this to over 110k training images across 501 scenes to support broader multi-view geometry problems.

Who it’s for

Researchers and developers working in 3D computer vision, specifically those focusing on multi-view stereo, camera pose estimation, novel view synthesis (such as NeRF and 3DGS), and depth/normal estimation.

Highlights

  • Large Scale: Includes 17k MVS training samples across 113 scenes in the base set, expanding to over 110k images in the BlendedMVG version.
  • Diverse Content: Covers a variety of categories including architecture and sculptures.
  • Multi-Resolution: Offers both low-resolution (768x576) and high-resolution (2048x1536) sets.
  • Broad Application: Useful for tasks like image retrieval, feature detection, and Transformer-based 3D vision.

Related

  • Project
    Tau-J/rtmlib

    rtmlib is a lightweight Python library for human and animal pose estimation that wraps RTMPose and ViTPose models. It avoids heavy OpenMMLab dependencies, requiring only numpy, OpenCV and ONNX Runtime (with optional GPU/accelerator back‑ends). The package provides high‑level solution classes (Wholebody, Body, Hand, Animal, Custom, Wholebody3d) that combine a detector (YOLOX, RFDETR, RTMDet) and a pose estimator, plus utilities for drawing skeletons. Models are auto‑downloaded from OpenMMLab or a HuggingFace mirror. Install via pip or from source, and run a Gradio WebUI or use the Python API for single‑image, video, or webcam inference.

  • Project
    rtr46/meikipop

    A universal Japanese OCR popup dictionary that enables instant word lookups from any on-screen content, including games, manga, and videos.

  • Project
    spectralpython/spectral

    A pure Python module for processing, visualizing, and classifying hyperspectral image data from imaging spectroscopy.

  • Project
    wenbowen123/BundleTrack

    BundleTrack is a real-time 6D pose tracking framework for novel objects that operates without requiring 3D CAD models of the object or its category.