최근 포토로그


Markerless Tracking Dataset Computer Vision

Scrapped from : http://www.metaio.com/research


Markerless Tracking Dataset

Overview
Unlikedense stereo, optical flow or multi-view stereo, template basedtracking lacks benchmark datasets allowing a fair comparison betweenstate-of-the-art algorithms. Until now, in order to evaluateobjectively and quantitatively the performance and the robustness oftemplate-based tracking algorithms, mainly synthetically generatedimage sequences were used. The evaluation is therefore oftenintrinsically based.

This website accompanies our ISMAR 2009 paper "A Dataset and Evaluation Methodology for Template-based Tracking Algorithms" (bib) in  whichwe describe the process we carried out to perform the acquisition ofreal scene image sequences with very precise and accurate ground truthposes using an industrial camera rigidly mounted on the end-effector ofa high-precision robotic measurement arm. For the acquisition, weconsidered most of the critical parameters that influence the trackingresults such as: the texture richness and the texture repeatability ofthe objects to be tracked, the camera motion and speed, and the changesof the object scale in the images and variations of the lightingconditions over time.
We designed an evaluation scheme for objectdetection and inter-frame tracking algorithms and used the imagesequences to apply this scheme to several state-of-the-art algorithms.The image sequences are freely available for testing, submitting andevaluating new template-based tracking algorithms.

 

How to use it
Belowyou find the datasets we generated until now. Each dataset consists ofa movie, an image of the tracking target, the intrinsics of the cameraused and a file giving ground truth positions for every 250th frame,all movies consist of 1200 frames each. There are five movies pertarget focusing on "Angle", "Range", "Fast Far", "Fast Close" and"Illumination". The movies are encoded with the lossless FFV1 codecfrom the ffmpeg-project (ffmpeg.org), a DirectShow codec is availableat http://ffdshow-tryout.sourceforge.net/. You can use e.g. Virtual Dub http://www.virtualdub.org/ to convert the sequences into still images if you need to.

Thetask now is to detect the target image in the frames of the movie. Allreference targets are 640x480 images. For every 250th frame, we providethe coordinates of four corners that are placed at the pixels (+- 512;+-384), the origin of the tracking target is in its middle (see imageon the right, the white frame represents the 640x480 px target, thereference points given for initialization lie on the diagonal). Allimages have their origin in the upper left corner.

Weoffer to evaluate the results you obtain with your tracking algorithmand send you the results. If you agree, we can additionally publishyour results on the webpage. To evaluate your results against the ground truth we have for every frame, please send an email to research(at)metaio.com where you attach a tabulator-separated log file of your experiments (1 per sequence) formatted like this example.

Weevaluate your log files and then send you the results (example resultsfor SIFT see below on the right). As measure we use the RMS of the fourpixels. A frame is considered successfully tracked if the RMS is below10 px.

For the evaluation results of SIFT, SURF, FERNS and ESM please refer to our paper.

Support:
This work was partially supported by BMBF grant Avilus / 01 IM08001 P.

 

Contact info
For comments and suggestions, feel free to contact research(at)metaio.com




덧글

댓글 입력 영역