change_log.md

Planned Features

• Add URDF loader to define articulated objects and robots
• Add support for more object datasets (ideas are welcomed, just open an issue with a dataset you want to see integrated)
• Support non-rigid physics simulation
• Improve performance
• Deprecate Global Storage

Version History

Version 2.2.0 17th December 2021

• switch blender version to 3.0.0 instead of 2.93.0
  • we now rely on Cycles X, making the rendering much faster than before
  • this also depreactes the usage of bproc.renderer.set_samples(), this is now replaced with bproc.renderer.set_noise_threshold(). This fcts allows to set the desired noise ration on a pixel basis, giving a much higher control to ensure a certain consistent noise level in the whole image.
  • it is still possible to limit the amount of samples per pixel with a new function named: bproc.renderer.set_max_amount_of_samples()
  • as the whole image is now rendered at once we removed the auto-tile addon
  • the BLENDER denoiser is no longer available, we recommend using the INTEL denoiser.
• remove the argument keep_using_base_color from bproc.lighting.light_surface() and Material.make_emissive(), now either a emission color or the base color is used
• changes for the bop integration
  • install bop_toolkit automatically
  • add BlenderProc2 python scripts for BOP challenge
  • allow to pass list of objects to write_bop for which to save pose annotations
  • divide load_bop() into load_bop_objs(), load_bop_scene() and load_bop_intrinsics() to decouple the functionalit
• fix a bug after uninstalling pip packages they were not truly removed

Version 2.1.0 17th November 2021

• add new lens distortion module, adding the possibility to simulate k1, k2, k3, p1 and p2 parameters.
• improve usability of distance rendering, make it the default in all examples
• distance and depth rendering can now be done antialiased and non-antialiased
• add a nocs renderer: paper
• upgrade the shell sampler and rework the usage and documentation
• move run.py to cli.py and support the same CLI arguments as for blenderproc ...
• make caching of pip installs the default (faster start-up) and add option to reset them if needed
• refactor the vishdf5 script using the same content now for saveAsImg
• clean up the scripts folder and move it completely into blenderproc
• include all packages (like vhacd) in the pip install package
• fix bug with light internal blender_obj after an undo operation
• add OPTIX as denoiser option
• add possibility to read jpeg
• fix use_ray_trace in bproc.sampler.upper_region
• added an exception if not the correct python environment is used

Version 2.0.0a7 13th October 2021

• fix Light class to work properly with bpy.types.Light in load_blend
• fix the stereo matching example add missing parameter to fct. call

Version 2.0.0a6 13th October 2021

• moved to a full python API support (.yaml files are still supported)
  • blenderproc can now be installed via pip: pip install blenderproc
  • complete new structure to most of the files
  • offer new command line interfaces: blenderproc run, blenderproc debug, blenderproc vis and many more
  • optional/flag arguments can now be used in the users python script
• adds feature to render multiple times in one session (see example advanced/multi_render)
• added tutorials to get easy start with BlenderProc
• adjusted all examples to explain how the python API would be used there
• restructured the documentation
• fixed a bug with the view layer update, which wasn't done before the scene coverage score was calculated, which caused problems in the 3D-FRONT examples.
• removed a lot of view layer update calls, which makes the whole execution faster, by manually calculation the matrix_world.
• vertex colors can now be used with a Principled BSDF shader
• all images are now moved into a seperate image/ directory
• all functions now use type hints

Version 1.12.0 23th July 2021

• switch to blender 2.93, with that textures are now stored on the GPU between different frames increasing the speed drastically
• moved a lot of modules to the python API
  • all camera samplers
  • all writer classes
  • sampler and getters
  • surface lighting
  • random room constructor
• all outer API functions work now with numpy arrays instead of mathutils classes
• regroup all examples to make them more easily accessible
• refactoring of the LabelIdMapping class
• sampler.Path now has a return_all and random_samples option
• add support for the new 3D front labels and script to find them more easily in the future
• AMASS now also supports torch=1.8.1+cu111, to better support modern GPUs
• refactor download script for haven
• added a flag to not center objects from the ShapeNet dataset
• added a platform independent unzipping function to solve the problems on windows
• add function to return the default_value or the connected node based on the input name of a BSDF shader node of a material
• fix for MacOS, where the blender version was not used in prior installations, leading to the problem that users had to remove blender on their own
• the writer now correctly writes the fov if asked

Version 1.11.1 18th May 2021

• fix a bug on mac os, where the blender python path has changed
• fix a bug in the BlenderLoaderModule which caused that the setting of custom properties was not successful

Version 1.11.0 11th May 2021

• BlenderProc now can be executed via a python script (experimental, not all modules are supported yet)
• python API extension for the following modules:
  • the Entity & Material Manipulation
  • the WriterUtility, Hdf5Writer, BopWriterUtility, CocoWriterUtility
  • the PostProcessing
  • the physic positioning, PoseSampler
• debugging can now be done via the command line, by just typing --debug after a BlenderProc call
• add automatic convex decomposition to speed up physics, object decompositions will be cached
• add a jupyter notebook example for google colab
• add a BasicEmptyInitializer module, which helps creating simple empty objects, these can be used for using a depth of field
• add a depth_of_field option to the CameraInterface (thanks to cuteday)
• add a hide module, so that certain objects can be hidden in a selected range of frames (thanks to Sainan Liu)
• add an option to change the computation type via the config
• blender 2.92 changed depth and distance once again, we use it correctly now and added an explanation
• added a diffuse rendering mode
• added a key frame API
• added a default config, which can be easily changed
• ObjectPoseSampler now can check also against only a limited set of objects
• ObjectReplacer now stops if no objects are left to replace
• SurfaceLighting now reuses the TextureLess material
• preloading in CCTexture now also works perfectly for textures which use an alpha texture
• python packages now are installed by specifying the required packages at the top of each file
• added an example for pasting objects onto random backgrounds
• added a flag to merge BlenderKit objects, while loading, they will be named after the blenderkit file
• reintroduce persistent transformation scaling to improve stability
• improve the clean up at the start of BlenderProc
• fixed a bug in the Front3DLoader, which avoided the correct setting of the custom properties
• fixed a bug with the once_for_all mode for the randomize materials fct.
• add a simple rerun script to show how BlenderProc is used for creating a diverse dataset
• switch to blender 2.92

Version 1.10.0 24th February 2021

• rewritten the BlendLoader, which is now able to only load objects of a given type, this will break existing config files as the load_from parameter has been replaced by datablocks and obj_types
• the MaterialManipulator can now overlay/mix a texture with a selected material
• the MaterialManipulator can now mix/add a material with another selected material, ideal for creating surface imperfections on other materials
• the MaterialManipulator can now add a layer of dust on materials, which simulates dust flakes on top of the object, there also exists now an example for it
• amount_of_repetitions can now be used with providers
• full integration of the blenderkit module, including modules and textures with an example
• LoaderInterface now offers an auto shading mode
• the CCMaterialLoader uses now a preselected list of probably useful materials by default to avoid non tileable materials
• restructure loaders and move their functionality to the utilities (necessary for the API changes for 2.0.0)

Version 1.9.0 10th February 2021

• all Loader now support setting the add_material_properties of the newly loaded objects
• each new .hdf5 file, will contain the git commit hash of BlenderProc
• introduce LightUtility -> this fulfills our long time goal giving BlenderProc a python API
• changed the default stereo mode from OFFAXIS to PARALLEL
• adapt to the changes made to 3D Front, as 3D Front does not have a version number, we support the newest version and a version from Summer 2020, if you have errors please open an issue
• getter.Material can now return all materials of a list of objects, selected via the getter.Entity
• add a cf_add_* fct to the MaterialManipulator, which works similar to the cf_set_* fct.
• add check_empty to getter.Entity and to getter.Material to throw an error if the returned list is empty
• added a cf_add_uv_mapping/forced_recalc_of_uv_maps option to force to recalculate the uv map of materials, which already have a uv map
• added option that randomized materials in cf_randomize_materials are added to objects without any materials
• change that cf_randomize_materials inside the EntityManipulator now deals with lists and not single elements as before
• add auto download for the imageio library, if that fails an exception is thrown with instructions on how to do this yourself
• CameraInterface now supports setting poses for a certain frame -> also needed for the python API
• changed the depth_scale default value in the BopWriter from 0.1 to 1.0, added an option to change via the config
• changed the ignore_dist_thres default value in the BopWriter from 5.0 to 100.0
• improved the ShapeNet example, by adding a ShapeNetWriter, which saves which object was used in the .hdf5 container
• fixes a bug with the separation of ceilings and floor objects in SceneNet scenes
• fix a bug, where the add_alpha_channel_to_texture fct. couldn't deal with empty material sockets
• fix the pyhsics_positioning example by using the collision_shape="MESH"
• fix a bug, where empty material_sockets would have been returned in the getter.Material
• fix vis_coco_annotation script, didn't work anymore after the segmentation map rewrite

Version 1.8.2: 27th January 2021

• added stereo mode to SegMapRenderer
• switched to using imageio for reading images, as the blender image loading API does not support .exr files written in stereo mode
• added option to write world-to-cam transformations to the BopWriter
• CocoAnnotationWriter does now in the polygon format not write empty segmentation lists anymore and uses iscrowd:0
• type hints are now added to the generated documentation
• extracted reusable rendering code from the renderer modules into utility classes

Version 1.8.1: 14th January 2021

• fixed a bug in the WriterInterface
• fixed cc_texture downloader script in the case of weird urls
• fixed creating the ceiling when there are rounding errors in the vertices coordinates

Version 1.8.0: 11th December 2020

• massively improve the documentation, by adding using github pages for our documentation: https://dlr-rm.github.io/BlenderProc/index.html
• add a RandomRoomConstructor:
  • This module can generate random floor plans of single rooms, with arbitrary extrusions.
  • It can also randomly place loaded objects without collision inside of the room
  • By using CCTexture it is also possible to assign random materials to the floor, wall and ceiling
• remove the SceneNetLighting module and replace it with the SurfaceLighting module, to make it more general
• add support for the haven websites and add an example and corresponding loaders:
  • The haven environment website: https://hdrihaven.com/
  • The haven model website: https://3dmodelhaven.com/
  • The haven texture website: https://texturehaven.com/
• add option to the CameraSampler to ensure that a certain object is always in the camera view
• improve the FloorExtractor to extract the floor and ceiling in SceneNet and other scenarios
• improve the skin tones for AMASSLoader
• add a proper scaling to the ikea dataset, by converting all files into SI units
• fix the ikea dataset downloading, by removing and splitting broken pieces, we advise to redownload the dataset

Version 1.7.0: 1st December 2020

• switch to blender 2.91
• added an example of how to set object poses via a transformation matrix and set camera extrinsics / intrinsics via a transformation matrix and a K matrix
• added camera utility class which makes it easier to set and retrieve intrinsics via any K matrix
• added loader for Pix3D dataset
• added loader for AMASS dataset
• added loader for the ikea dataset
• added community driven support for Windows
• added motion blur and rolling shutter support
• fixed collision checks between meshes, so the ObjectPoseSampler is not generating colliding object poses anymore
• fixed cleanup of temporary directories in the case of an error
• set pixel origin to the center of a pixel, slightly affecting intrinsics and depth outputs
• fixed coco annotations if background is not visible
• added support for transparent background
• fixed wrong image size in coco annotations
• fixed blender proc if non-english language is configured
• when loading ply files a default material is now added
• fixed setting matrix_world via the entity manipulator
• fixed the scenenet examples (corrects physics positioning and camera sampling)

Version 1.6.1: 25th August 2020

• fixed bbox computation in the Coco Annotations
• fixed visualization of Coco Annotations
• add set properties fct. to BlenderLoader
• adds a script to download all materials from BlenderKit (not used yet)

Version 1.6.0: 24th July 2020

• added an TextureLoader, which can load images and store them as blender textures
• added a provider to access these loaded textures
• removed the MaterialRandomizer
• moved the whole functionality from the MaterialRandomizer to the EntityManipulator
• fixed a bug in the SUNCG material loading, where for a few objects the diffuse material was confused with an alpha material
• add set_properties for Front3DLoader
• fixed the min_interest_score that it works with more than just SUNCG
• fix a bug in the visHdf5 scripts with distance maps which had several channels
• improved documentation for min_interest_score and check_pose_novelty_translation

Version 1.5.1: 17th July 2020

• adding support for 3D-Front: https://tianchi.aliyun.com/specials/promotion/alibaba-3d-scene-dataset
• added to example on how to use 3D-Front
• fixed visualization of newly generated SegMapRenderer results
• fixed combining several coco annotations
• added option for the "class" key for default values in the SegMapRenderer
• fixed a bug in the SegMapRenderer when used in debug mode

Version 1.5.0: 14th July 2020

• rename depth to distance (this is the difference between taking at each pixel the Z coordinate (in the camera coordinate system) or the distance from the optical center, as sensors usually produce depth and not distance, we changed this)
• create postprocessing module to calculate the depth out of the distance image
• rewriting of the SegMapRenderer, it now supports the mapping of any attribute or custom property of an object to an image or a csv file
• reworking of the used coco examples to better work with this new SegMapRenderer
• adapting CocoAnnotationsWriter to write category_ids which must be defined in a Loader or in a .blend file
• adding support for a supercategory inside of the coco annotations tool (useful for bop dataset names)
• add texture provider, which can generate random textures, which can be used in a displacement modifier
• added to the EntityManipulator:
  • option to automatically generate a UV mapping for an object
  • added option for displacement modifier adding to objects
  • added an example for the entity_displacement_modifier usage
• the sampler.Value now supports gaussian sampling
• reworked some parts of the Hdf5Writer, to better work with depth and distance
• added a check to the CocoUtility to avoid that the bounding box has a area of zero
• remove the provider.ContentGetter as it was confusing, now it works directly without it
• add an Interface in the naming of some modules, if they are not supposed to be run by the pipeline directly
• change the sample amount for the FlowRenderer to 1 as a higher sampling density integrates over several objects in one pixel
• added a step function in the alpha mode of the renderer to avoid artifacts, when rendering semantic segmentations
• support for setting different fx/fy values in the camera module
• remove the pixel_aspect_x/pixel_aspect_y/resolution_x/resolution_y from the renderer is now in Camera
• fixed a bug when the unknown texture folder in the scenenet dataset was not created before
• improved the loading of camera parameters from files
• added a feature for supporting changes of the "category_id" for the WorldManipulator
• improved the visHdf5/saveHdf5/generate_nice_vis scripts
• switches to blender 2.83.2

Version 1.4.1: 3rd June 2020

• switches to blender 2.83
• adds support for adaptive sampling (Renderer/use_adaptive_sampling) -> can decrease rendering time for complex scenes if used
• add random_samples option for entity getter
• add friction/damping for physics positioning
• add JPG save option for RgbRenderer
• add uniform elevation sampling in Shell Sampler
• add in plane rotation option for look_at/direction camera rotations
• add postprocessing.Dist2Depth to convert Blender distance images to depth images
• UpperRegionSamlper can sample in relative area on selected face
• improve physics by using convex hull and box collision shapes
• obj.rigid_body.use_margin = True, to actually use margin
• add a Smooth Shader cf to entityManipulator
• BopWriter improvements, outputs chunks of 1000 images
• exclude objects like planes from proximity checks
• fix transparency bounces
• add custom property cp_bop_dataset_name
• add bop_challenge example with all configs used to render the provided synthetic data
• add bop_object_on_surface_sampling for sampling upright objects
• docu improvements and step by step explanations of config files

Version 1.4.0: 15th May 2020

• added the NormalRenderer functions to the RGBRenderer saves a render call, when generating normal images
• added a script to automatically download all PBR assest from https://cc0textures.com/
• add CCMaterialLoader, which can load all downloaded materials from cc0textures, to randomly replace them later
• add support for ShapeNet
• add support for SceneNet, including the lighting, which is modeled after the SUNCG lighting module.
• massive docu improvements, added a type to all config values and a default value if one is there
• added examples for:
  • bop_object_physics_positioning
  • bop_object_pose_sampling
  • bop_scene_replication
  • on_surface_object_sampling
  • scenenet
  • scenenet_with_cctextures
  • shapenet
  • shapenet_with_scenenet
  • shapenet_with_suncg
  • suncg_with_improved_mat
• added object OnSurfaceSampler (similiar to the phyiscs but faster and places object only on the bb)
• added WorldManipulator (can change custom properties of the world, like the category_id)
• switched to config version 3.0
  • global initialized values are now in the main.Initializer
  • added a GlobalStorage to save variables over module boundaries
• add VisNormalImage for debugging mode, which visualizes a depth & normal image in blender
• material selectors now also support AND and OR conditions
• add automatic detection of optix graphics cards
• added a min interest score for the CameraSampler
• added a novelty pose checker for the CameraSampler, to avoid clustering of camera poses
• global config values are now not longer added to the modules config, only in the event that there is no module found the global config is checked
• rename SUNCG materials to the textures they load, makes selecting them easier
• SUNCG materials are now shared over objects, if they share the same properties, also true for lights
• rewrite of MaterialRandomizer to make use of the new cc0textures
• added a Bop Writer to write out the current scene config in the BOP format
• add the solidify modifier to the options of the EntityManipulator
• add a PartSphere sampler to e.g. only sample in the upper half of a sphere
• add a script to download scenenet to make the use of it even easier
• add easy option to change specular values for materials with the MaterialManipulator
• added the option to make runs deterministic by introducing a env variable: BLENDER_PROC_RANDOM_SEED
• added a Basic Color Sampler
• improved the capabilities of the DiscSampler
• now all custom properties should be accessed in the config via: "cp_..." for custom functions: "cf_"
• add texture support for ycbv objects in the BopLoader
• change the default results of the SegMapRenderer to NYU ids instead of the row-index in the .csv file.

Version 1.3.0: 25th March 2020

• added Optical Flow Renderer
• added Stereo Global Matching Renderer (SGM) -> which takes two color images and produces "non-perfect" depth images
• added option for SGM to also write disparities
• added a new material manipulator
• added a material getter to select materials based on certain conditions
• combined BOP with this repo -> now the only difference is the readme
• added contribution guidelines -> which include commit message guidelines and branch name guidelines and much more
• matrices can now be read via the config file
• added a cached options to the import objects
• added request custom property for the entity getter (to ensure that a custom property is changed)
• added ReplaceObject
• added option that in debugging the RgbRenderer get executed except for the actual rendering part all others are executed and then undon
• added collision mesh source to the physics module
• added option to coco annotations to append the output to an existing coco file
• added bounding box selection to getter.Entity and also just axis aligned hyper plane checks
• added a 1D sampler for float/int/bool
• added postprocessing to the example to reduce the amount of channels in the depth image
• added visualization for stereo and optical flow
• added an oil filter for the SGM result
• added a path sampler
• added a blender collection loader
• added the RockEssentialLoader based on the RockEssential dataset
• added index to getter.Entity to request only a certain element of the selection
• fix three bugs in the physics module, where the mass_scaling was always used and the location not properly updated
• fixed a bug where color images where wrongly saved in float now back to uint8
• fixed a bug in the coco annotations writer + adapting of the example
• fixed bug in load_image, where the dimensions where switched
• fixed a bug when "home_local" is not available than "home" is used
• fixed bug where run.py can only be run from the main folder
• fixed pypng version in bop example
• fixed bug where the volume calculation for bounding boxes did not work as expected
• refactored the disc sampler
• cleaned up all readmes

Version 1.2.0: 31th January 2020

• added more detailed examples to most of the important modules
• adapt CoCo annotation tools to newer Version of SegMapRenderer
• background class is now zero when doing instance segmentation
• renamed getter.Object to getter.Entity, same for ObjectManipulator is now named EntityManipulator
• entity conditions do now work with AND and OR connections
• strings are now matched with fullmatch instead of search in entity conditions
• conditions now support bool custom properties
• material randomizer now supports getter.entity providers
• added UpperRegionSampler, which can sample on the up surface of the bounding box of an object
• scripts now support execution with python3.x
• physics are now saved with bool instead of active and passive
• physics options for simulator added
• remove version number -> only blender version 2.81 supported
• the ObjectLoader can now load several objects at once and set their properties with "add_properties"
• moved more functionality in the camera sampler instead of having in specific sampler to make writing a new one easier
• added a SO3 Sampler for rotation sampling
• change the "name" of modules in the config to "module" and "provider" depending on the case
• added MacOS support (but only for CPUs, GPU support on MacOS is not available)
• added this changelog
• fixes:
  • fixed a problem that the check_bb_intersection did not work right in all cases (Bounding box check)
  • fixed a problem with the check_intersection fct., where the reference coordinate system was not always right.
  • fixed a bug with the SUNCG lighting in RGB images

Version 1.1.0: 16th January 2020

• Added provider, which can be executed inside of different modules, examples:
  • Samplers of values
  • Getters of objects, specified on one condition (for example the name)
• Added object manipulators, these can change selected objects attributes and custom properties
• Added physics positioning, objects can be either "active" or "passive" and then interact with each other
• Added ObjectPoseSampler sampled collision free poses of all objects
• Added one ObjectLoader, which deals with all kinds of objects
• Added a camera object sampler to sample cameras and objects at the same time
• Redone the SegMapRenderer to fix bug in instance and class segmentation
• Added coco annotations writer
• Added version number to config files
• Added more documentation
• Added more examples

Version 1.0.0: 25th October 2019

• Added Pipeline, Modules and Config
• Added CameraModules, with different sampler for SUNCG and replica
• Added LightModules, with different samplers also for SUNCG
• Added Loader for objects
• Added Loader for replica and SUNCG
• Added Hdf5 Writer
• Added camera-, light-, object state writer
• Added MaterialRandomizer
• Added Shell-, Sphere- and Uniform3DSampler
• Added debug script