SiSter: Single camera Stereo Robot

Original paper: Beyond the Baseline: 3D Reconstructionof Tiny Objects with Single CameraStereo Robot

Table of contents

• Dataset
• C++ Implementation
• ROS Implementation

Sister Dataset

Download dataset

One Drive Link (n.b. click on "Download" on the top menu to download a single zip file).

Description

The dataset contains the following objects:

arduino
component_0J
component_1B
component_1G
hexa_nut
hexa_screw
nodemcu
washer

Each subfolder image has this naming convention:

<OBJECT_NAME> / <CAMERA_DISTANCE> / <BASELINE> / <center|left|top|right|bottom>.png

for example:

component_1G/10cm/025mm/left.png

means the left view of the rig with a baseline of 25mm with a camera-to-object distance of 10cm, of the component with name component_1G.

Each groundtruth depth image, instead, has a similar naming convention:

<OBJECT_NAME> / <CAMERA_DISTANCE> / gt_depth.exr

given that the depth groundtruth is associated with the center image of each rig and it varies with the camera distance (but not with the rig baseline).
The EXR file stores the depth in meters as a float32 array. To read it with python and opencv you can use:

import cv2
depth = cv2.imread('gt_depth.exr', cv2.IMREAD_ANYCOLOR | cv2.IMREAD_ANYDEPTH)

Sister C++: 5 views disparity computation

This repository containes a prototype implementation of the multiview stereo algorithm described in the paper. The implementations is contained in the cpp folder and is basically wrapped around a single helper class.

Requirements

OpenCV 
Eiegen3
OpenMP

Compile

cd cpp
mkdir build
cd build
cmake ..
make

Run example

The above compile command will build the library along with a sample application compute_disp that will load 5 images and calculate multiview disparity. You can run the example on the sister dataset folders themselves. For example you can run

./compute_disp $SISTER_DATASET/component_1G/10cm/025mm/ 192

where $SISTER_DATASET is the root folder of the dataset described above. The second argument is the full path of the folder containing the 5 images (pay attention to naming convention, images names should be <center|left|top|right|bottom>.png). The second argument 192 is the max disparity value.

ROS implementation

ROS Example