hbb2obb-analyze¶
Breaks an evaluation down by orientation, size, frame edge, difficulty and class. Guide: Analyze an evaluation.
usage: hbb2obb-analyze [-h] [--version] [--hbb_dir HBB_DIR] [--img_source IMG_SOURCE]
[--map_path MAP_PATH]
[--excluded_classes EXCLUDED_CLASSES [EXCLUDED_CLASSES ...]]
[--boundary_classes BOUNDARY_CLASSES [BOUNDARY_CLASSES ...]]
[--iou_threshold IOU_THRESHOLD] [--class_agnostic] [--out_dir OUT_DIR]
[--no_bar]
gt_dir pred_dir
Break down OBB conversion accuracy by ground-truth orientation, size, edge, difficulty and class, against the horizontal prompt as a baseline
positional arguments:
gt_dir Directory containing ground truth OBB annotations
pred_dir Directory containing predicted/converted OBB annotations
options:
-h, --help show this help message and exit
--version, -V show program's version number and exit
--hbb_dir HBB_DIR, -hd HBB_DIR
Directory of the HBB prompts, to score the conversion against emitting
them unchanged (default: gt_dir/../labels_hbb, if it exists)
--img_source IMG_SOURCE, -i IMG_SOURCE
Image or directory of images, for the frame-edge cut and for relative HBB
coordinates
--map_path MAP_PATH, -mp MAP_PATH
Path to label map YAML file (optional)
--excluded_classes EXCLUDED_CLASSES [EXCLUDED_CLASSES ...], -e EXCLUDED_CLASSES [EXCLUDED_CLASSES ...]
Class IDs to exclude
--boundary_classes BOUNDARY_CLASSES [BOUNDARY_CLASSES ...], -bc BOUNDARY_CLASSES [BOUNDARY_CLASSES ...]
Class IDs the side-ratio table is restricted to (default: all)
--iou_threshold IOU_THRESHOLD, -t IOU_THRESHOLD
IoU threshold for considering a match (default: 0.1)
--class_agnostic, -ca
Match boxes regardless of class labels
--out_dir OUT_DIR, -o OUT_DIR
Write analysis.md, analysis.yaml and analysis.png here as well as printing
the report
--no_bar, -nb Disable tqdm progress bar display
The hbb2obb toolkit ships seven commands:
hbb2obb convert HBB annotations to OBBs by prompting SAM
hbb2obb-detect detect HBBs with an Ultralytics model, for images that have none
hbb2obb-convert convert annotations between YOLO, DOTA, Pascal VOC, COCO and LabelMe
hbb2obb-view inspect HBB and OBB annotations over their images, interactively
hbb2obb-eval score predicted OBBs against ground truth (IoU, orientation error, per class)
hbb2obb-analyze break that score down by orientation, size, edge, difficulty and class
hbb2obb-optimize search hyperparameters, as one sweep or a whole benchmark
Run any of them with --help for its own options, e.g. hbb2obb-detect --help.