v1.0.0

bbox-visualizer is now stable. The API in this release is the long-term shape of the library: future 1.x releases will be bug fixes only, and any breaking change would mean a 2.0.

Breaking changes

• All functions are now pure. Every public function returns a new image and never modifies the input. Previously add_label, add_T_label, and draw_flag_with_label drew on the input array in place. If you relied on in-place mutation, keep the return value instead: img = bbv.add_label(img, label, bbox).
• suppress_warnings and warnings_suppressed removed. Fallback warnings now go through Python's standard logging module. To silence them:

```python import logging

logging.getLogger("bbox_visualizer").setLevel(logging.ERROR) ```

• Python 3.10 is now the minimum supported version (3.9 is EOL).

New

• Support for COCO and YOLO bounding box formats. All drawing functions accept a bbox_format keyword argument, so you can pass boxes straight from your dataset or model without converting them yourself:

python img = bbv.draw_rectangle(img, [0.5, 0.5, 0.2, 0.3], bbox_format="yolo")

Supported formats: "voc" ([x_min, y_min, x_max, y_max]), "coco" ([x_min, y_min, width, height]), and "yolo" (normalized [x_center, y_center, width, height]). VOC remains the default, so existing code is unaffected.

• draw_multiple_boxes / draw_multiple_rectangles accept a sequence of colors, one per box: bbox_color=[(0, 255, 0), (0, 0, 255)]. A single tuple still applies to all boxes.
• Runnable example scripts in examples/.

Fixed

• Rectangle strokes are drawn inward so their outer edge lies exactly on the bbox coordinates, aligning boxes with labels (previously the stroke was centered on the coordinates and spilled outside the box).

• The multi-object functions now raise ValueError when labels and bboxes lengths differ, instead of silently ignoring extras.

• Label backgrounds are sized from the measured ink of the rendered text (including flag and T-labels): larger size / thickness values no longer overflow the background, and padding is even on all sides instead of bottom-heavy.

• The flag pole no longer pokes above the flag background, and it rises at least 50px above the box (matching the T-label stem) so it stays visible on small boxes.
• add_T_label and draw_flag_with_label fallback paths now forward size, thickness, draw_bg, text_bg_color, and text_color to add_label instead of dropping them.
• The multi-object functions accept numpy arrays of bounding boxes (previously raised ValueError: The truth value of an array is ambiguous).
• Colors may contain numpy integer channels, e.g. sampled straight from image pixels (tuple(img[y, x])); they are normalized to built-in ints before reaching OpenCV.
• Labels are placed above the box only when the full label background fits in the frame, so small labels near the top edge fall back inside the box instead of being clipped.

Security

• The minimum opencv-python version is now 4.8.1.78 (was 4.1.0.25). Older opencv-python wheels bundle a libwebp vulnerable to CVE-2023-4863, a critical heap buffer overflow triggerable by decoding a malicious WebP image. The new floor prevents installs from resolving those builds. If you already run opencv-python 4.8.1.78 or newer (any release since September 2023), nothing changes.
• The release pipeline was hardened following a security audit: all GitHub Actions are pinned to full commit SHAs, the publish workflow no longer interpolates the tag name into shell (removing a script-injection vector), and workflow token permissions are scoped to the jobs that need them. Publishing continues to use PyPI trusted publishing (OIDC), with no long-lived API tokens.

Internal

• CI consolidated into a single workflow running make check (ruff + ty + pytest) on Python 3.10, 3.12, and 3.14.
• Dependabot security alerts and automated security updates are enabled on the repository; there are deliberately no scheduled version-bump PRs.