View annotations

hbb2obb-view draws annotations over the images they belong to, in a window that pans and zooms, or into image files:

hbb2obb-view data/images                               # converted OBBs over their HBBs
hbb2obb-view data/images --show_confidence             # OBBs colored by confidence, score printed
hbb2obb-view data/images --compare data/labels_obb_gt  # ground truth overlaid in blue
hbb2obb-view data/images -o annotated                  # write annotated images, open no window

What it reads

img_source is one image or a directory of them. Each layer comes from a directory beside it unless a flag names one:

Layer	Flag	Default, the first that exists
OBBs	--obb_dir / -od	../labels_obb, ../labels_obb_gt, ../labels
HBBs	--hbb_dir / -hd	../labels_hbb
segmentation polygons	--polygon_dir / -pd	../labels_polygon
confidence side-car	--confidence_dir / -cd	../labels_confidence
a second OBB set	--compare / -c	none

So hbb2obb-view images opens a converted set, a ground-truth set or a plain images/ + labels/ release without further flags. Class names come from -mp, or from a classes.yaml or names.txt beside the labels or one level up.

Colors

Color	Layer
green	the OBBs; colored green to red by confidence with --show_confidence or the c key
white	the HBBs
red	the segmentation polygons (written by hbb2obb --save_polygon)
orange	boxes flagged difficult
blue	the --compare set

The confidence comes from a 10th column on the OBB lines or from the side-car directory, which is read whether or not --show_confidence was given, so c works either way.

Formats

The viewer reads every format hbb2obb-convert writes. When a directory holds the same boxes in several formats, it reads the YOLO files by default, and also offers a COCO file paired with the directory (coco_annotations_<name>.json beside labels_<name>/). It prints what it found:

OBB: yolo from data/labels_obb, also dota, coco
HBB: yolo from data/labels_hbb, also voc, coco

In the window, t and y step the OBB and HBB layers through those formats, re-reading every frame, and the status bar names the one on screen. --obb_format and --hbb_format pin one at startup. --obb_dir and --hbb_dir also accept a COCO .json file directly.

Keys

Key	Action	Key	Action
q, Esc	quit	o	show or hide the OBBs
n, →, ↓	next frame	h	show or hide the HBBs
p, ←, ↑	previous frame	l	show or hide the class labels
wheel, +, -	zoom about the cursor	d	show or hide boxes flagged difficult
drag	pan	c	color the OBBs by confidence, and print it
f, 0	fit the frame to the window	g	show or hide the segmentation polygons
1	zoom to 100%	t / y	read the OBBs / HBBs from the next format
s	save the current view to <frame>.view.jpg in the working directory	x	cycle the overlay: both sets, primary only, --compare only

--frame / -f opens a given frame first, by file stem. --window sets the window size (default 1600x900). --hide_hbb and --hide_class_labels start with those layers off.

Writing images

--out_dir / -o draws every frame at full resolution to <out_dir>/<frame>.jpg instead of opening a window; --frame restricts it to one.

--crops writes a contact sheet of the individual objects rather than whole frames, as <frame>.crops<n>.jpg, to --out_dir or the working directory. --indices restricts the sheet to some boxes, by their row in the label file, to review particular objects:

hbb2obb-view data/images --crops --frame 04433 --indices 0 1 2 -o review

All flags: hbb2obb-view reference.