cenas

 

Introductory summary

 

The objective of this SHREC track is to retrieve 3D models captured with a commodity low-cost depth scanner. This dataset, is intended to be an extension of the work done for SHREC'13, but now using the new Kinect One, which provides more detail. The fast the spread of such cameras, has increase the need of better and more detailed datasets. Considering the characteristics of this context, we captured another two hundred more, common household objects. These range from cups and dishes, to staplers, ash trays and so on.

 

Our dataset provides up to 90 frame pairs of RGB and Depth images for each object, its corresponding registered and segmented point cloud and a polygon mesh model of all objects, carefully processed, instead of just collections of local views. As can be seen by the sample presented in the picture above, these models are still rough representations of models. Although there are already exists techniques that can create more detailed meshes of captured point-clouds, with our dataset we intent to provide a dataset of meshes that could be used in real-time capture and recognition.

 

 

Task description

 

In response to a given set of queries, the task is to evaluate similarity scores with the target models and return an ordered ranked list along with the similarity scores for each query. The query set is a subset of the larger collection.

 

 

Ranked lists


Each file representing an evaluation should be named %id-%query where %id is the number of the run and %query is the identifier of the object representing the query. Please submit the results on any ASCII file format such as '.res'.

Example (for a faux query 144.off):

144    1.00000
24      0.87221
45      0.79915
201    0.59102
203    0.54902
32      0.51241
...(etc)...


This would output in a file named, for instance 1-144.res

 

 

Data set

 

The collection is composed of 192 scanned models, which were acquired through the real-time capture of 224 collected objects. Of these, 32 were rejected due to low quality or material incompability. The range images are captured using a Microsoft Kinect One camera. The collection is presented in three different ASCII file formats: PLY, OFF and STL, representing the scans in a triangular mesh.

 

The collection itself is uncategorized. The rankings are to be compared against a human-generated ground truth created with a group of 20 subjects.

 

A test set is available here: here

 

 

Evaluation Methodology

 

We will employ the following evaluation measures: Precision-Recall curve; E-Measure; Discounted Cumulative Gain; Nearest Neighbor, First-Tier (Tier1) and Second-Tier (Tier2).

 

 

Procedure

 

The following list is a description of the activities:

 

 

 

Schedule

 

To be updated, will start in the 2nd week of January
February 7 - A test set will be available on line. This will be a subset of the final set.
February 11 - Please register before this date.
February 11 - Distribution of the final query sets. Participants can start the retrieval.
February 18 - Submission of results (ranked lists) and a one page description of their method(s).
February 20 - Distribution of relevance judgments (Human-generated) and evaluation scores.
February 25 - Submission of final descriptions (two page) for the contest proceedings.
February 28 - Track is finished, and results are ready for inclusion in a track report
March 29 - Camera ready track papers submitted for printing
May 11 - 8th EUROGRAPHICS Workshop on 3D Object Retrieval including SHREC'2015

 

 

 

  Any further questions can be sent to shrec@3dorus.ist.utl.pt