[Dataset] the MIT LabelMe

[Scene] Street;  Collected from 17 Folders (Table 3)

[dataset selection] Table 1

[6 object classes] building, tree, sky, car, road, body (person).

[Image preprocessing] downscale to less than 400*300 while keeping the aspect ratio.

[Accuracy] Table 2; (the classification performance of human body, cars and trees is NOT good).

[Testing results]  [download 24M] (up left: input image; up right: ground truth; down left: boosting result; down right: probability map of ground truth;)

[Parameter settings] For Cascaded boosting: 15 layers; 800 weak classifiers per layer; DetectRate = 0.998; False Positive Rate = 0.7;

[Reference] Long Zhu, Yuanhao Chen, Yuan Lin, Chenxi Lin, Alan Yuille: Recursive Segmentation and Recognition Templates for 2D Parsing. NIPS 2008.

Table 1: Dataset

86.0

Table 2: Testing Results.

Global  Accuracy :     86.0

Average Accuracy :     70.1

Confusion Matrix: 
    90.3 building:    90.3     2.8     1.6     1.5     3.5     0.3	
    66.1 tree    :    29.9    66.1     1.5     1.3     1.0     0.2	
    88.6 sky     :     8.2     2.1    88.6     0.1     0.8     0.1	
    69.4 car     :    19.1     1.9     0.0    69.4     8.6     0.9	
    90.8 road    :     6.2     0.5     0.0     2.1    90.8     0.4	
    15.7 body    :    60.2     1.2     0.0     8.6    14.3    15.7	
Avg:70.1

Table 3: 17 Folders containing street scene from the MIT LabelMe dataset; http://labelme.csail.mit.edu/