3D Rigid Tracking from RGB Images Dataset

The dataset challenge consists in tracking 3 rigid, poorly textured, highly occluded objects across sequences of monocular RGB images.

For each object dataset, we provide:

  • a simple CAD model of the object (.obj)
  • several learning videos
  • one or more testing videos
  • the groundtruth pose of the camera wrt the object reference system for all the learning videos
  • a simple Matlab script Test.m showing how to employ the pose groundtruth.
All videos were shot using a CANON EOS 5D camera, f = 50 mm.
THIS IS NOT A RGB-D dataset. No depth image is provided for any of the datasets.
The goal consists in retrieving the pose of the camera wrt to the object in all the frames of the testing videos, For training your algorithms, you are allowed to use anything but the testing video frames.


This dataset contains 5 learning videos and 2 testing videos showing an electric box filled and emptied with several objects.


Ths dataset contains 4 learning videos and 2 testing videos showing a non-textured food can over 2 different backgrounds; the can is occasionally occluded by a user’s hand grasping it and several distractor objects that create clutter.


This dataset contains 8 learning videos and 1 testing video showing a non-textured office door moving over a cluttered background. On the testing sequence, a user opens the door and passes through it.


Additionally, we provide accurate manual annotations of some parts of the target objects on the testing sequences. These can be used to test 2D detectors for localizing 3D objects undergoing perspective and light changes.


Part of the manual annotations were kindly provided by Mahdi Rad of the Graz Univertisy of Technology.


All the video sequences (except the BOX negative video) are provided as 1920 x 1080 png images. The file XXX-TestAndInfo contains the testing video sequences of each dataset, the 3D model (.obj) a READ_ME, a python script for importing the pose data and a simple matlab example.
Clic on the corresponding link to download the data.
BOX Dataset

Negative video          BOX-Learn-video1       BOX-Learn-video2

BOX-Learn-video3    BOX-Learn-video4       BOX-Learn-video5

CAN Dataset

CAN-Learn-Part1    CAN-Learn-Part2    CAN-TestAndInfo

DOOR Dataset

DOOR-Learn-Part1    DOOR-Learn-Part2    DOOR-TestAndInfo

2D Annotations

2D annotations


The 3D pose estimation can be evaluated computing the L2 norm of the rotation and translation components of the absolute pose error [2] for all the frame of each video sequence, and evaluating their Cumulative Distribution Function (CDF) .

Quantitative results are provided in [1] in the form of the normalized Area Under Curve (AUC) score for each error. The AUC score is computed dividing the area of the CDF curve by the max error of the graph. The max error was set to 0.5 for both rotation and translation for all frames. See [1] for further details.


You are free to download and use these data for research purposes only.
We appreciate an email message indicating who has copied the data.
By downloading and using the dataset you agree to acknowledge its source (CVLab EPFL) and to cite the paper [1] in case results obtained with these data are published.
For any further details, questions and remarks, you can write to:
alberto [dot] crivellaro [at] epfl.ch


A. Crivellaro; M. Rad; Y. Verdie; K. M. Yi; P. Fua et al. : A Novel Representation of Parts for Accurate 3D Object Detection and Tracking in Monocular Images. 2015. International Conference on Computer Vision (ICCV), Santiago, Chile, December 13-16, 2015.

[2] Sturm, J., Engelhard, N., Endres, F., Burgard, W., & Cremers, D. A benchmark for the evaluation of RGB-D SLAM systems. IROS 2012