Learning Object Intrinsics from a Single Internet Image

Home

Discover

Bookmarks

Keep Up With Latest Trending Papers. Computer Science, AI and Machine Learning and more.Subscribe

CC BY

Source arXiv

Computer Vision

Learning Object Intrinsics from a Single Internet Image

Seeing a Rose in Five Thousand Ways

We build a generative model that learns to capture object intrinsics (distribution of geometry, texture, and material) for a wide range of objects, each from a single image, such as a photo of a bouquet.

Our method achieves superior results on multiple downstream tasks, including intrinsic image decomposition, shape and image generation, view synthesis, and relighting.

Authors

Yunzhi Zhang, Shangzhe Wu, Noah Snavely, Jiajun Wu

Related Topics

Generative model

Adversarial framework

Random sampling

Neural representation

3d-aware generative models

Rendering engine

Generative framework

Colmap @ cite1

Multi-view reconstruction or neural rendering methods

Multi-view reconstruction methods

Introduction

We present a model for shape and image generation under novel viewpoints and illumination conditions, as illustrated in.

Specifically, our model takes the single input image and learns a neural representation of the distribution over 3-d shape, surface albedo, and shininess of the object, factoring out pose and lighting variations, based on a set of instance masks and a given pose distribution of the instances.

The resulting model enables a range of applications.

Our contributions are three-fold: we propose the problem of recovering object intrinsics, including both 3-d geometry, texture, and material properties, from just a single image of a few instances with instance masks.

We design a generative framework that effectively learns such object intrinsics without overfitting the limited observations from only a single image.

Result

We propose a method that recovers the object intrinsics (distributions of geometry, texture, and material, separated from extrinsics such as poses and lighting) from a single image containing multiple instances of the same object type with masks.

We have developed a neural representation to model such intrinsics and an adversarial framework for training.

The proposed method successfully recovers object intrinsics on various objects from internet images, enabling many applications including shape and image generation, novel view synthesis, and relighting.

Read the Paper

Read Paper

Bookmarks

Bookmarks

Keep Up With Latest Trending Papers. Computer Science, AI and Machine Learning and more.Subscribe

CC BY

Source arXiv

Computer Vision

Learning Object Intrinsics from a Single Internet Image

Seeing a Rose in Five Thousand Ways

Our method achieves superior results on multiple downstream tasks, including intrinsic image decomposition, shape and image generation, view synthesis, and relighting.

Authors

Yunzhi Zhang, Shangzhe Wu, Noah Snavely, Jiajun Wu

Related Topics

Generative model

Adversarial framework

Random sampling

Neural representation

3d-aware generative models

Rendering engine

Generative framework

Colmap @ cite1

Multi-view reconstruction or neural rendering methods

Multi-view reconstruction methods

Introduction

We present a model for shape and image generation under novel viewpoints and illumination conditions, as illustrated in.

The resulting model enables a range of applications.

We design a generative framework that effectively learns such object intrinsics without overfitting the limited observations from only a single image.

Result

We have developed a neural representation to model such intrinsics and an adversarial framework for training.

Read the Paper

Read Paper

Home

Discover

Bookmarks

◐ Recommended Follows

◐ Latest Activity