CVPR`16 results

The following files are the CVT sequences and correspondences from the algorithm presented in the paper "Volumetric 3D Tracking by Detection," at CVPR`16.
The recording was done in Kinovis 4D modelling platform and CVTs are computed using the method here.

If you use our results, please cite our paper:

Volumetric 3D Tracking by Detection
C.-H. Huang, B. Allain, J.-S. Franco, N. Navab, S. Ilic, E. Boyer
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, June 2016.

Ballet:
	template: Ballet.zip (13.4 MB) seq1: rawCVT (81 MB); deformed templates (27 MB) seq2: rawCVT (66 MB); deformed templates (63 MB) associations from forest: corres (19 MB)
Thomas:
	template: Thomas.zip (22.1 MB) seq1: rawCVT (114 MB); deformed templates (83 MB) seq2: rawCVT (115 MB); deformed templates (76 MB) associations from forest: corres (35 MB)
Goalkeeper:
	template: Goalkeeper.zip (13 MB) sideJump: rawCVT (25 MB); deformed templates (26 MB) upJump: rawCVT (18 MB); deformed templates (41 MB) associations from forest: corres (4 MB)

Notice:

Deformed templates are obtained from [Allain et al. CVPR`15], where we have temporal-coherent centroid indices. We consider them as ground truths to train with (seq1) and to test with (seq2).

Forest predictions can be found in each corres. link. Each row represents the prediction for each input cnetroid in the deformed templates.
The third column means the matched point index on the template (visu_color4matching.off). If the predictions are 100% correct, this column should look like 0, 1, ..., 4999.

See here for parsing the association files and this matlab demo code to see how I compute error. You also need this to load the template.

We also emulate a more pratical tracking experimental setting, where input is noisy raw CVTs. Please refer to Sect. 5.2 in the paper for more explanation. Forest predictions are NOT provided in this case.