Bachelorarbeit

Automatic dataset generation for neural depth estimation in active setups

Requirements

• python version 3.9.18
• BlenderProc
• torch version matching cuda version eg. for ACID: conda install pytorch=1.11.0 torchvision=0.12 torchaudio=0.11 cudatoolkit=11.5 -c pytorch -c conda-forge
• python3 -m pip install -r req.txt
• Download Resources and unpack folder to ./blenderproc/

Usage

Dataset generation and evaluation pipeline

usage: pipeline.py [-h] [--loop | --no-loop] [--output_dir OUTPUT_DIR] [--num_poses NUM_POSES] [--num_images NUM_IMAGES] [--training | --no-training] [--load_scene LOAD_SCENE] [--load_camera LOAD_CAMERA]
                   [--save_scene SAVE_SCENE] [--save_camera SAVE_CAMERA] [--save_camera_intrinsics SAVE_CAMERA_INTRINSICS] [--model {Unet,Unetresnet,all}] [--combined COMBINED]
                   [--pattern {rainbow,continuous,points,stripes}]
----
pipeline from random-room-generation to neuronal monocular depth-estimation
----
optional arguments:                                                                                                                                                                                             
  -h, --help            show this help message and exit
  --loop, --no-loop     enable/disable image-generation (default: False)
  --output_dir OUTPUT_DIR, -o OUTPUT_DIR
                        name of outputfolder (default: meow_stacked)
  --num_poses NUM_POSES, -poses NUM_POSES
                        Number of poses within one frame (default: 5)
  --num_images NUM_IMAGES, -images NUM_IMAGES
                        Number of created frames (default: 250)
  --training, --no-training
                        (un)set trainings loop for depthestimation (default: True)
  --load_scene LOAD_SCENE
                        File to load scene from (default: )
  --load_camera LOAD_CAMERA
                        File to load camerapose from (default: )
  --save_scene SAVE_SCENE
                        File/folder to save scene to (default: )
  --save_camera SAVE_CAMERA
                        File/folder to save camerapose to (default: )
  --save_camera_intrinsics SAVE_CAMERA_INTRINSICS
                        File to save camera-intrinsics from/to (default: )
  --model {Unet,Unetresnet,all}
                        select model type (default: all)
  --combined COMBINED   Stack Normals and Infrareds (default: True)
  --pattern {rainbow,continuous,points,stripes}
                        select pattern type (default: points)

Visualize hdf5 data

The default value for the maximum depth is set to 20. Since a normalization in the range [0,1] (scientific standard) was performed, this value must now be adjusted.
blenderproc vis --depth_max=1 hdf5 <path/to/file.hdf5>