Dataset Version 0.1 - 15.06.2014

Admin Luca Fiaschi - luca.fiaschi@gmail.com

References:

[1] @ARTICLE{ fiaschi_13_keeping,
	 year = { 2013 }, 
	 pages = { 656-659 }, 
	 journal = { ISBI 2013.Proceedings }, 
	 author = { L. Fiaschi and  K. Gregor and  B. Afonso and  M. Zlatic and  F. A. Hamprecht }, 
	 title = { {Keeping Count: Leveraging Temporal Context to Count Heavily Overlapping
	Objects} }, 
	 timestamp = { 2013.01.23 }, 
	 doi = { 10.1109/ISBI.2013.6556560 }, 
	 cite = { fiaschi_13_keeping  } 
}

[2] @INPROCEEDINGS{ fiaschi_14_tracking,
	 year = { 2014 }, 
	 author = { L. Fiaschi and  F. Diego and  K. Gregor and  M. Schiegg and  U. K{\"o}the and  M. Zlatic and  F. A. Hamprecht }, 
	 title = { {Tracking indistinguishable translucent objects over time using weakly
	supervised structured learning} }, 
	 booktitle = { CVPR. Proceedings, in press }, 
	 timestamp = { 2014.03.20 }, 
	 cite = { fiaschi_14_tracking  } 
}

If using this dataset please cite reference [2].

We provide the original data used in the experiments presented in [2].
Dependencies: the data is stored as hdf5 files, please refer http://www.h5py.org/ for specifications.

Various populations of 72 hours old Drosophila larvae were filmed for 5 minutes with a temporal resolution of 3.3 frames per second, 1000 frames in total. Images have a spatial resolution of 135.3 μm/pixel, a size of 1400 × 1400 pixels. For detection and segmentation of the foreground regions we use the open-source software ILASTIK (www.ilastik.org). After elimination of tiny isolated objects (τ < 15 pixels), we compute connected components of each thresholded foreground probability image of the series. The graph G is created by linking foreground regions from neighboring timesteps that overlap spatially, by more than 10 pixels. All foreground regions which are not fully inside a margin of 100 pixels from the image borders are excluded to avoid dealing with truncated larvae (cluster).

This folder contains:

1) data - contains all the movies in the dataset [2] e.g. name_movie = t8_2012_10_18_1543_data4
    1.1) data/name_movie  contains the intermediate results of algorithm [1] and the raw data
    1.1.1) data/name_movie/name_movie.h5  hdf5 file containing the data
            * The hdf5 file format is t,z,x,y,c 
            * The volume "volume/data" contains the volume of the vide 8bit
            * The volume "volume/pmap" contains the probability map for the background/foreground segmentation
            * The volume "volume/segmentation" contains a foreground/ background segmentation
            * The volume "volume/segmentation" contains detected connected components of the foreground, each connected
              component has assigned a unique integer, 0 indicates background
            
    1.1.2) data/name_movie/name_movie.graph_fast : contains the graph information from algorithm [1]
            * The array "graph_fast/entire/graph_fast" is compressed version of the graph edges.
              eg. [ 1, 13, -1,  2, 14, -1,  3, 15, -1,  4 ... ] indicates edges between connected components labeled 
                    1-13 , 2-14, 3-15, ... where -1 is a stopper value.
            * The array 3*N "graph_fast/entire/positions" contains the position of the pixels of the connected components. eg [ [1 1 1], [  0 686 529], [  0 686 530], [  0 686 531], [  0 686 532], [  0 686 533], [  0 689 549], [  0 689 550], [-1,-1,-1], [2,2,2] ]. The first row is the connected component label while [-1,-1,-1] is a stopper value. Values between [1,1,1] and [-1,-1,-1] are pixels positions as t,x,y
            * The array "graph_fast/entire/sizes" is array indicating the size (in pixel) of the connected components.
              e.g. array[1] is the size in pixel of the connected component labeled 1
            * The array "graph_fast/entire/times" is the time of which connected component is labeled.
              e.g. array[1] is the time where connected component 1 is present.
    1.1.3) data/name_movie/name_movie.res : contains the inferred number of objects per connected component from algorithm [1].
            * The array "graph_fast/entire/true_sizes" contains the number of objects per connected component.
              e.g. array[1] is the number of objects in connected components 1.
    1.1.4) data/name_movie/movie_sizes : contains a visualization of the results of algorithm [1]. Please refer to the original paper.
    1.1.5) data/name_movie/movie_gt_tracks : contains a visualization of the GT tracks. This values were not used in the evaluation of [2] which is only based on the overlapping spatiotemporal regions. Tracks shorter than 20 should be filtered out as could be due to spurious errors in the manual tracking.
    
    1.1.6) data/name_movie/name_movie.txt : contains the GT tracks for the movie.

2) cvpr_all_enconters: overlapped spatiotemporal regions are obtained by parsing the graph of algorithm [1] and used in the evaluation of algorithm [2]. These are divide per encounter type.

    Encounter file format specifications
    
    * The hdf5 volume format is t,z,x,y,c 
    * The volume subgroup contains the cropped information for the overlappend spatiotemporal region.
    * The volume "volume/userbrush" is the ground truth labelling from the user marking each connected component before entering the overlapping region and after exiting the overlapping region with a unique brush stroke. Integer value.
    * The array "meta/info" indicates the location of the spatiotemporal region in the original movie volume. For example [info[0]:info[1]+1 ] indicates the region in t, [info[2]:info[3]+1 ] idicates the region in x and [info[4]:info[5]+1 ] indicates the region in y.
    * The array "meta/inlarvae" contains the id of the connected components of isolated larvae entering the region.
    * The array "meta/outlarvae" contains the id of the connected components of isolated larvae leaving the region.