I had a little free time on my hand and decided to quickly complete the coursera-course „Data Science at Scale – Practical predictive analytics“ of the University of Washington by Bill Howe. The last assignment was to participate in a kaggle competition.
For this assignment I chose the „San Francisco Crime Classification“ challenge. The task is to predict the Category of a crime given the time and location. The dataset contains incidents from the SFPD Crime Incident Reporting system from 2003 to 2015 (878049 datapoints for training) with the following variables:
In the video below you can see the sequence of a car driving in a city scene and braking. The layers I rendered out for groundtruth data are the rendered image with the boundingbox of the car (top left), the emission layer ( shows the brakelights when they start to emit light, top right ), the optical flow (lower left), and the depth of each pixel in the world scene ( lower right).
Render-time was about 10h on a Nvidia GeForce GTX 680, tilesize 256×256, total image-size: 960×720.
Setting up groundtruth rendering, saving and reading it again
After you have composed your scene, switch to the Cycles rendering engine, and enable the render-passes in the properties-view (under render-layers) you plan on saving as ground-truth information. In this case, I selected the Z-layer, which represents the depth of the scene:
Now open the Node-editor-view, display the Compositing-Nodetree and put a checkmark at „Use Nodes“, it should look like this:
Hit „T“, and in the „output“-tab, select „File Output“. Connect it to the Z-output of the render-layers-node. Hit „N“ while the File-Output-node is selected, enter an appropriate File-subpath as name for the data (I chose „Depth“ here), and depending on your usecase, change the output format. As I am expecting float-values that are not constrained to 0-255, I will save the data in the OpenEXR-format:
After rendering ([F12]), we have a file called /tmp/Depth0001.exr. It can be read and displayed using python:
author: Tobias Weis
import numpy as np
import matplotlib.pyplot as plt
def exr2numpy(exr, maxvalue=1.,normalize=True):
""" converts 1-channel exr-data to 2D numpy arrays """
file = OpenEXR.InputFile(exr)
# Compute the size
dw = file.header()['dataWindow']
sz = (dw.max.x - dw.min.x + 1, dw.max.y - dw.min.y + 1)
# Read the three color channels as 32-bit floats
FLOAT = Imath.PixelType(Imath.PixelType.FLOAT)
(R) = [array.array('f', file.channel(Chan, FLOAT)).tolist() for Chan in ("R") ]
# create numpy 2D-array
img = np.zeros((sz,sz,3), np.float64)
data = np.array(R)
data[data > maxvalue] = maxvalue
data /= np.max(data)
img = np.array(data).reshape(img.shape,-1)
depth_data = exr2numpy("Depth0001.exr", maxvalue=15, normalize=False)
fig = plt.figure()
The output of this script on the default-scene looks like this:
Further layers to generate groundtruth data
Besides depth information, we can also generate groundtruth data for other modalities:
Object/Material index: Can be used to generate pixel-accurate annotations of objects for semantic segmentation tasks, or simply computing bounding boxes for IOU measures
Normal: Can be used to generate pixel-accurate groundtruth data of surface normals
Vector: Will show up as „Speed“ in the node editor, contains information for optical flow between two frames in a sequence
Emit: Pixel-accurate information of emssion-properties.
What is the target of all this ? Driving in an automotive scenario with a given speed and turnrate at any moment, we want to predict the displacement of a 2D-projection (pixel) between two frames:
By using the camera-calibration, I can create artificial curves and walls as 3D point-sets and project them back to 2D. Using discretized values for speed, turnrate, streetwidth and wall-height, I can then simulate the displacement of these 3D-Points when they are projected to 2D (our image).
(Note for me: this is the backprojection-code, main-file: main_displacements.py)