A dockerised service for Bounding Box and Segementation data generation using Blenderproc

Requirements:

Linux 20.04
Python 3.5.2+
Docker
Kubernetes

Setup

Prepare objects/scenes

Objects/scenes are available across the web. A couple of good free sites:

Objects: Free3d
Scenes: PolyHaven

Prepare output folder structure

The microservice by default outputs data to /data/nfs1/synth, using /data/nfs1/synth/temp during creation. Before commencing, create this directory structure and set permissions to -rwxrwsrwx using:

chmod -R 2777 /data/nfs1

The Data Generation process

This microservice creates a kubernetes pod/docker container which fulfils requests. The process:

Uses Swagger CodeGen to generate server code
Builds the BlenderProc docker
Builds and deploys the data generation microservice
Sends a data generation request after a small wait to test the service end to end
Further requests can be sent by modifying the request.json and running:

make curl_docker

Annotations are produced in the standard COCO json format.

Run the Data Generation microservice in docker

make all_docker

Run the Data Generation microservice in kubernetes

make all_kube # if sudo is required
make all_nosudo # if not

NB:

The kubernetes version expects a NFS to be configured, with its ip address specified in synth-pv-pvc.yaml: (spec:nfs:server)

Modifying the scene

The request.json specifies details of how the scene is produced. The flow of the program is as follows:

A point of interest is specified in the scene where the model is placed
The program stochasticly shifts the point of interest (within a range) so the object isn't in the direct centre of the screen (camera_shift_range_xyz)
The program specifies a random camera location within a range (camera_location_range_xyz)
The rotation matrix calculates the camera angle change required to point the camera towards the point of interest

rowdentech/BlenderProc-service