The annotation framework can be used to generate pseudo labels from the model prediction and handle this data for manual human inspection. This includes merging and splitting images and finally convert them into the hdf5 file format.
The expected file naming contention is key_index.suffix, with a 6 digit integer for index. This is required to process the files properly. See reindexing to automatically convert your data into the right format.
The pseudo label generation module can be used to convert the predicted heatmaps into json label files. It provides the following CLI interface.
usage: generate_pseudo_label.py [-h] [-s SUFFIX] [-r RES] [-bt BIN_THRESHOLD]
[-md MIN_DIAMETER]
input_dir output_dir
Generate pseudo labels from predicted heatmaps
positional arguments:
input_dir Path to the predicted heatmaps
output_dir Path to the RGB images where the generated labels
belong to and should be stored
optional arguments:
-h, --help show this help message and exit
-s SUFFIX, --suffix SUFFIX
Suffix of the image files. (default: .png)
-r RES, --res RES Single camera resolution WIDTHxHEIGHT (default:
640x480)
-bt BIN_THRESHOLD, --bin_threshold BIN_THRESHOLD
Values over this threshold will be binarized to 1
(default: 96)
-md MIN_DIAMETER, --min_diameter MIN_DIAMETER
Minimum diameter for an ROI in percent to the image
width (default: 0.05)The model cannot ensure to detect all ROIs consistently for all views. To simplify the labeling process, use this module which syncs the ROI for all views based on a given camera layout
usage: create_roi_consistency.py [-h] [-o OUTPUT_DIR] [-c CAMERA_CONFIG] [-f FOV_DEGREE] [-i IOU_THRESHOLD] input_dir
Create ROI consistent label json files for multiple views
positional arguments:
input_dir Path to the directory which contains the images and labels in json format.
optional arguments:
-h, --help show this help message and exit
-o OUTPUT_DIR, --output_dir OUTPUT_DIR
Path to the directory where the roi consistent label files will be put. (default: _out)
-c CAMERA_CONFIG, --camera-config CAMERA_CONFIG
Path to the camera config file that contains the camera positions (default: record/config/6_camera_setup.ini)
-f FOV_DEGREE, --fov-degree FOV_DEGREE
Field of camera view in degree (default: 90.0)
-i IOU_THRESHOLD, --iou-threshold IOU_THRESHOLD
IOU threshold to adjust if a new ROI circle need to be added (default: 0.7)The merge module can be used to combine individual frames and label files into a merged frame, to inspect all camera views at the same time. It provides the following CLI interface.
usage: merge.py [-h] [-o OUTPUT_DIR] [-s SUFFIX] [-r RES]
[--image_topics IMAGE_TOPICS [IMAGE_TOPICS ...]]
[--images_per_row IMAGES_PER_ROW] [--hdf5] [--reindex]
input_dir
Merge images and json labels
positional arguments:
input_dir Path to the directory which contains the images and
labels in json format.
optional arguments:
-h, --help show this help message and exit
-o OUTPUT_DIR, --output_dir OUTPUT_DIR
Path to the directory where the generated files will
be put. (default: annotation)
-s SUFFIX, --suffix SUFFIX
Suffix of the image files. (default: .png)
-r RES, --res RES Single camera resolution WIDTHxHEIGHT (default:
640x480)
--image_topics IMAGE_TOPICS [IMAGE_TOPICS ...]
All image topics that should be should be merged
together. The order defines the layout of merging
(default: ['front_left', 'front', 'front_right',
'rear_left', 'rear', 'rear_right'])
--images_per_row IMAGES_PER_ROW
Number of images that are aligned next to each other
(default: 3)
--hdf5 Merge files into hdf5 file (default: False)
--reindex Reindex image and label files to a sequential
continuous numbering (default: False)Additionally, it provides the following 2 features.
Run with --reindex to reindex the data using the correct format. It will delete frames that do not have an image for every view as we assume this means an identical timestamp. The other images will be shifted to a sequential indexing starting at 000000.
Images generated with the recording module are not aligned and need to be reindexed before further usage.
Images and labels can be combined into a HDF5 file by running with --hdf5.
The split module can be used to split the merged images and label files back into individual ones. By default splitting the images is disabled to safe memory and speed up the runtime as usually the individual images before merging are still available. The script provides the following CLI interface.
usage: split.py [-h] [-o OUTPUT_DIR] [-s SUFFIX] [--split_images] input_dir
Split images and json labels
positional arguments:
input_dir Path to the directory which contains the images and
labels in json format.
optional arguments:
-h, --help show this help message and exit
-o OUTPUT_DIR, --output_dir OUTPUT_DIR
Path to the directory where the generated files will
be put. (default: annotation)
-s SUFFIX, --suffix SUFFIX
Suffix of the image files. (default: .png)
--split_images Split merged images into individual ones. This is
disabled by default safe memory and speed up the
runtime as usually the individual images before
merging are still available. (default: False)
Use h5_extract.py to extract data from one or more hdf5 files.
usage: h5_extract.py [-h] [-o OUTPUT_DIR] h5_files [h5_files ...]
Extract data from hdf5 file
positional arguments:
h5_files Path to the hdf5 files.
optional arguments:
-h, --help show this help message and exit
-o OUTPUT_DIR, --output_dir OUTPUT_DIR
Path to the output directory (default: annotation)