1. liang_chieh_chen
  2. segKITTI

Overview

HTTPS SSH

Beat the MTurkers: Automatic Image Labeling from Weak 3D Supervision

This website stores the dataset and codes used in the paper by Liang-Chieh Chen, Sanja Fidler, Alan Yuille and Raquel Urtasun, CVPR 2014

For now, only some parts of the dataset and codes are released. The demo code will be provided soon.

My Email: lcchen@cs.ucla.edu

Folders

  • code : codes used in our experiments

We refer a car as a tracklet in our experiments.

  • AnnotMasksBD_left : stores the bounding box posistions (of each tracklet) for the annotated images. This is provided by our in-house annotators.

  • AnnotMasks_left : stores the annotated masks for the images. (For visualization, load the image, and use imagesc in MATLAB). This is provided by our in-house annotators.

  • Images_left : stores all the images used in our experiments (those images are randomly selected from raw folder in KITTI dataset)

  • tracklets : stores all needed data for segmenting tracklets.

Folders under tracklets:

  • tracklet_depth : For each tracklet, the depth values of each point (saved in sparse binary format)

  • tracklet_gt : ground truth of car segmentations provided by the nine in-house annotators

  • tracklet_cadShape : car shapes generated by CAD models.

  • tracklet_mtShape_batch1 : car shapes provided by MTurkers batch 1

  • tracklet_mtShape_batch2 : car shapes provided by MTurkers batch 2

  • tracklet_mtShape_batch3 : car shapes provided by MTurkers batch 3

  • tracklet_pc : projected point clouds for each tracklet

    • White : used as foreground seeds.
    • Black/Dark Grey : background seeds.
    • Grey : missing points.
  • tracklet_recon_depth_GauMRF : reconstructed depth values from point clouds using Gaussian MRF (saved in dense binary format)

  • tracklet_recon_depth_LapMRF : reconstructed depth values from point clouds using Laplacian MRF (saved in dense binary format)

  • tracklet_roi : images for each tracklet

  • tracklet_star_pc : topological cue used in our experiments, i.e., the star shape prior.

    • xx_pe.bin:saved in sparse binary format,
    • xx_pi.bin:saved in dense "int" binary format, this is the node indices for each edge
  • tracklet_stereo_depth : reconstructed depth values from stereo (saved in dense binary format)

  • model_param : trained model parameters

See the function load_all_data in code/opencv/source/Utility.cpp for reference.

dense binary format:

Basically, the first two fields (int) are row, and column size of the data matrix, followed by "float" data scanned in the row-order.

sparse binary format:

The first two fields (int) are row, column of the data matrix, followd by tuples of (r, c, fval), where r is the row position, c is the column position, and fval is the float value for each tuple.