Skip to content

Latest commit

 

History

History
86 lines (70 loc) · 3.6 KB

quickstart.md

File metadata and controls

86 lines (70 loc) · 3.6 KB
Raw

Quick Start Guide

This guide is designed to get a classification network training for your dataset ASAP. If you need something more specific read through the rest of these readme files and feel free to ask on Gitter Join the chat at https://gitter.im/Fastcode/VisualMesh.

This guide assumes you're using docker and have already installed it.

Setting up

To setup with docker once you have docker installed all you need to do is build the docker image.

docker build . --pull -t visualmesh:latest

Making a dataset

Gathering the data

Create a folder that contains your source training data. You will need three files per image:

imageX.jpg maskX.png
Image Mask

Lens.yaml

projection: EQUISOLID
focal_length: 420
centre: [0, 0]
k: [0, 0]
fov: 1.80248
Hoc:
  - [-0.244225, -0.905877, 0.346036, 0]
  - [-0.0304556, 0.363831, 0.930967, 0]
  - [-0.96924, 0.216827, -0.116446, 0.8]
  - [0, 0, 0, 1]
  • Hoc is a homogenous transformation matrix which if used like Hoc * (a,b,c,1) would convert a vector in the cameras coordinate system into one from the ground. In this matrix, only the z component of translation matters (the height of the camera above the ground). In this example, the units are measured in meters (m). However, the unit that you use for this measurement don't matter, so long as the units you use for the radius of your geometry are the same.
┌            ┐
│          0 │
│    R     0 │
│          z │ <- Only Z is important for the visual mesh
│ 0  0  0  1 │
└            ┘

  • projection is either RECTILINEAR, EQUISOLID or EQUIDISTANT depending on the lens used in the camera.

  • focal_length is the focal length of the camera measured in pixels.

  • centre is the offset from the centre of the lens, to the centre of the image in pixels. If you don't know set it to [0, 0]

  • k is the two component radial distortion values for the lens as per a polynomial series. If you don't know set it to [0, 0]

  • fov is the field of view of the camera, for many rectilinear lenses it will be the full diagonal field of view, but for lenses with black circles (like above) it will be the field of view of that black circle.

Building the dataset

Once you have these in a folder make sure they are named imageX.jpg, maskX.jpg and lensX.yaml where X is replaced with a unique number per example.

You can then run

./training/build_dataset.py <path/to/images> <path/to/output>

and it will create a training, validation and testing TFRecord file containing the data.

Training

In order to train, you must create a configuration file. To start with you can copy example_net.yaml. You must place this configuration file in the output folder under the name config.yaml. For example, copy over example_net.yaml to output/quickstart/v1/config.yaml

Running the following command will then begin training the network.

./docker run --gpus all -u $(id -u):$(id -g) -it --rm --volume $(pwd):/workspace visualmesh:latest ./mesh.py train <path/to/output>

Testing

Once your network has trained you are able to run the following command to test the network.

./docker run --gpus all -u $(id -u):$(id -g) -it --rm --volume $(pwd):/workspace visualmesh:latest ./mesh.py test <path/to/output>

It will measure several properties of the network and output them to the output directory.