Desk3D Dataset:

This webpage details the Desk3D dataset for depth-based object instance recognition. If you use this dataset please cite [1].

The dataset contains 6 object instances: Face, Ferrari, Kettle, Mini, Phone, Statue

The test scenes are obtained by performing fusion [2] on 5 consecutive frames. Based on the testing challenge we divided the test scenes into 4 main scenarios:

Test scenario 1:
  • Contains challenge to detect two similar objects (Mini and Ferrari) in clutter.
  • Has 3 scenes each with 40 fused frames.
Test scenario 2:
  •  Contains challenge to detect single object with large pose variation.
  •  Has 5 scenes (one for each object excluding Ferrari) each with 100 fused frames.
Test scenario 3:
  • Contains challenge to detect multiple objects in large clutter.
  • Has 6 scenes each with 100 fused frames.

Test scenario 4:
  • Contains challenge to detect multiple objects in occlusion and clutter. 
  • Has 5 scenes with 100 fused frames each.

Further we also have a test scene with only background to test the number of false positives and determine the classifiers separability while plotting precision-recall curves.

Download files:
There are 7 main download files compressed with tar.gz:

readMe.txt: A download version of this text.
train.tar.gz: Contains the training models of each instance along with training sequences of random background.
testScenario1.tar.gz: Contains the test scenes of test scenario 1.
testScenario2.tar.gz: Contains the test scenes of test scenario 2.
testScenario3.tar.gz: Contains the test scenes of test scenario 3.
testScenario4.tar.gz: Contains the test scenes of test scenario 4.
testClutter.tar.gz: Contains test scenes with only background clutter.

In addition, we also have a "codes.tar.gz" file that contains *.m files to read the data and a oni.tar.gz file that contains all the video sequence captured using a kinect which contains both the RGB and depth fields. The oni.tar.gz is the largest file (~8GB) and will be made available only on request. You'll need [3] to read from the raw oni files.

Description of Files:

For each object we have 8 different scans of the object saved as pose*/cloud_mesh_"object".ply.
Each scan is accompanied with a "tfrom.aln" file which can be used to transform them into their canonical pose.
The ground-truth of test scenes is with this canonical pose.
In addition we also give a mesh model saved as "mesh.ply" for each object.
Their are 3 background training sequences: scene_clutter1 (100 frames), scne_clutter2 (50 frames) and scene_clutter3 (100 frames).
In [1], scene_clutter1 & 3 were used for training and 2 for validation.

TestScenario1-4 and testClutter.tar.gz:

For each test scene, the fused frames are save as fusion/cloud_*ply.
The ground truth for each object is saved as dataIcp_"object"_check.mat.
These files contain three main fields:

  • tempExist: is 1 if the given object exists in the frame else 0.
  • vld: is 1 if the object is correctly ground-truthed and not cut off by borders else 0.
  • frmM: contains a 4x4 matrix for each frame giving the ground truth position.
In [1] test frames where both tempExist and vld are 1 were used for testing of the given object.

[1] Bonde, U., Badrinarayanan, V., & Cipolla, R. Robust Instance Recognition in Presence of Occlusion and Clutter. In ECCV 2014.
[2] R. Rusu and S. Cousins. 3D is here:  Point Cloud Library (PCL). In ICRA 2011