NTL-Remi / ApolloScope

ApolloScape dataset loading tool

🔎 ApolloScope - ApolloScape dataset loading tool

WIP - Version 0

This is a Python package for visualisation and loading of Baidu's ApolloScape data-set for vehicle vision tasks.

Currently it is only focused on the Scene parsing data-set.

Usage for v0.1.0

As of v0.1.0, this project is still a bit messy. I plan on cleaning it and improving the UI (probably inducing some heavy changes) when I will have some time. Here is some code if you want to use it in its current state:

import apolloscope
from apolloscope.ls_sp.register import Register, SequenceId, TypeId

Specify the paths to the data-set:

sp_path = PATH_TO_YOUR_SCENE_PARSING_FOLDER
ls_path = PATH_TO_YOUR_LANE_SEGMENTATION_FOLDER

apolloscope.root_folder.scene_parsing(sp_path)
apolloscope.root_folder.lane_segmentation(ls_path)

The expected folder architecture is the one used in ApolloScape archive files. Some archives seem to have been zipped differently than others and will not recreate the top-most folder (the one that has the same name as the archive), you will have to create it yourself (this is for example the case for the scene parsing road02_seg_depth.tar.gz file).

This will parse the file paths in the specified folders and classifying them by image type and sequence in a multi-indexed data-frame (see it like a 2D array in which we are going to slice depending on the data we want):

register = Register()

Say that we want to iterate at the same time on the colour data, the semantic segmentation and the depth maps. We define the three data types:

image_type = TypeId(dataset="SP",
                    section="seg",
                    subsection="ColorImage",
                    file_type="jpg")
depth_type = TypeId("SP", "seg_depth", "Depth", "png")
seg_type = TypeId("SP", "seg", "Label", "bin.png")

Suppose that we want to iterate over the frames captured on road 2, sequence 22, camera 5. We define the sequence:

test_sequence = SequenceId(road=2, record=22, camera=5)

In both data types and sequences definition, all parameters are optional, allowing to select larger parts of the data-set. Defining

test_sequence = SequenceId(road=2, camera=5)

will take the data of all records on road 2 filmed by camera 5.

To actually select the data from register, we do:

filtered_register = register.types([file_type, depth_type, seg_type])
filtered_register = filtered_register.sequences([test_sequence])

We can then get the Pytorch dataset:

dataset = apolloscope.ls_sp.pytorch.Dataset(filtered_register)

and use it in a pytorch dataloader in a classical way. In the current case, each element would be a tuple of three images corresponding to the three data types we defined, at a same time-stamp. Triplets with missing data would be dropped.