SUNRGB-D 3D Object Detection Challenge

Introduction

3D object detection is a fundamental task for scene understanding. In this task, we focus on predicting a 3D bounding box in real world dimension to include an object at its full extent. The test data consist of 2860 newly acquired RGB-D images that ground-truth bounding boxes are not publically available. We use the existing SUNRGB-D dataset as training data. This challenge is hosted with LSUN challenge in CVPR.

Data

• Training Set: SUN RGB-D dataset (10,355 RGB-D scene images) [SUNRGBD.zip] and [Groundtruth]
  
• Testing Set: 2860 newly collected RGB-D images [SUNRGBDLSUNTest.zip] and their [Metadata]
• Code to load the data and groundtruth data: [SUNRGB-D toolbox]

Evaluation and submission:

We evaluate 3D object detection by extending the standard evaluation criteria for 2D object detection to 3D. Assuming the box aligns with the gravity direction, we use the 3D intersection over union of the predicted and ground truth boxes for 3D evaluation with 0.25 as the IoU threshold. We use mean average precision as final evaluation metric. Please submit a downloadable link to Shuran Song the result file.

Submission format:

• There are in total 19 object classes. For each object class there will be one Matlab ‘.mat’ files to save the result. You can report results of only subset of these 19 object classes. For each ‘.mat’ result should be 1xN cell array, where N is number of test images.
• In each cell save an array of detection boxes of the corresponding object category and test image. It can be empty when there is no detection boxes in the image.
• For each detection box we need following fields, with same meaning as the groundtruth boxes in the SUNRGB-D toolbox:
  basis: 3x3 matrix
  coeffs: 1x3 matrix
  centroid: 1x3 matrix
  conf: double