dataset_document.md

Dataset Document

Introduction

The proposed OnePose_LowTexture dataset contains 40 objects with 80 sequences, and eight objects are coupled with a scanned model. The main part of the dataset is placed here, and scanned models are placed here. In our experiment, we use all of the objects for the test. Note that scanned models are only used for evaluation and to facilitate further research. Our method does not require a known object model.

Data Structure

The data structure of OnePose_LowTexture is the same as the OnePose dataset:

|--- lowtexture_test_data
|       |--- id-objname-category
|               |--- box3d_corners.txt
|               |--- objname-1
|                       |--- Frames.m4v
|                       |--- intrinsics.txt
|                       |--- color
|                       |--- intrin_ba
|                       |--- poses_ba
|                       |--- reproj_box
|               |--- objname-2

There are multiple sequences for each object. We use the first sequence (objname-1) for reconstruction and the last sequence (objname-2) for evaluation, similar to OnePose. For each object:

• Frames.m4v is the captured object video.
• intrinsics.txt contains an intrinsic matrix of original images in the video. All of the original images share the same intrinsic.
• box3d_corners.txt saves eight corners' coordinates of annotated object 3D bounding box.
• color directory contains all of the cropped foreground images (resized to $512\times512$). For each cropped image i.png, its corresponding intrinsic file and pose file are located in intrin_ba/i.txt and poses_ba/i.txt, respectively.
• The intrinsic file in the dataset contains the $3\times3$ projection matrix of the corresponding image. And the pose is defined as a $4\times4$ homogeneous transformation from the object system to the camera coordinate system.