1. Semantic Dataset Format

1.1 Raw Image Sets

    The raw image sets contain RGB images in PNG format, each of which is in the resolution of 500 (height) * 1000 (width).
    The raw image set "ParallelEye_rgb_train.rar" and corresponding semantic segmentation image set "ParallelEye_semantic_train.rar" can be downloaded from the homepage. In semantic segmentation task, the semantic segmentation images are provided only for the images in folders "04", "05" and "06".

1.2 Ground Truth Sets

    For the raw image sets, the ground truth of each training image is provided also as the RGB image in PNG format. Each pixel in ground truth image is labeled with RGB color, representing 10 semantic classes and a void class, according to Table 1.

Table 1: Color definitions of the semantic classes

Index Class R G B
1 Sky 128 128 128
2 Vegetation 128 128 0
3 Building 128 0 0
4 Fence 64 64 128
5 Sidewalk 0 0 192
6 Road 128 64 128
7 Traffic sign 192 128 128
8 Traffic light 0 64 64
9 Pole 192 192 128
10 Car 64 0 128
11 Void 0 0 0

    Figure 1 shows an example of semantic segmentation data. The raw image is shown on the left and the pixel-level semantic segmentation ground truth is shown on the right.

Figure 1. Example of semantic segmentation data. Left: raw image. Right: semantic segmentation ground truth.

2. Segmentation Task

    For each test image, you should segment all pixels into 10 semantic classes.
    The output from your computer vision system should be a PNG image with the pixels colored according to Table 1.

3. Evaluation

3.1 Data supplied

    The test set ("segmentation&depth_testing_set.rar" in homepage) contains 1022 RGB images in the same PNG format with training images.

3.2 Submission of Results

    The test results should be submitted as collections of PNG image files. One PNG image file should be generated for each test image and named in the same way, with pixel values ranging from (0,0,0) to (255,255,255). Some test images may include only a subset of the 10 semantic classes, in which case the results will be evaluated on only the included classes.

3.3 Evaluation method

    Each segmentation competition will be evaluated by average segmentation accuracy across the 10 semantic classes. Note that pixels labeled as void do not contribute to the evaluation. The segmentation accuracy for each class is assessed using the intersection-over-union metric, defined as the number of correctly labeled pixels of that class, divided by the number of pixels labeled with that class in either the ground truth labeling or the inferred labeling. Specifically, the segmentation accuracy for each class is given by the following equation:

    Participants are expected to submit a single set of test results for your team.