# Release Notes

{% hint style="success" %}
Read [how to update](https://doc.celantur.com/container/requirements-and-installation/updates) to the latest version of Celantur Container.
{% endhint %}

## 2026

### Version 26.04.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Released on April 3rd, 2026.

* Upgrade software dependencies in Docker image to reduce vulnerabilities.

### Version 26.03.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Released on March 16th, 2026.

* Improve segmentation mask processing. User can remove disconnected parts (controllable via environmental variable.)

### Version 26.02.2 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Released on February 24h, 2026.

* Improve EXIF metadata copying in JPEG data.
* Improve logging
  * for telemetry (if enabled)
  * for video processing (more verbose error message if post-processing fails)
* Manual garbage collection and GPU cache clearing to improve memory performance and reduce risks of memory overflow.
* FIX issue in REST API mode, where the DELETE request does not delete the metadata JSON of video files in the output folder within the container.
* FIX Issue in  REST API mode, when processing a defect video,  `status`  is "done" instead of "failed".
* \[EXPERIMENTAL] Enable gathering of hardware utilisation metrics with `--monitor data.json`

### Version 26.02.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Released on February 4h, 2026.

* New scheme for [#instance-segmentation](https://doc.celantur.com/usage/segmentation-masks-and-metadata#instance-segmentation "mention").
* Adjust error handling in licensing verification.

### Version 26.01.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Released on January 29th, 2026.

* Expose object tracking parameters as environmental variables.
* Customer can disable video post-processing, where audio and data streams from the input video are merged into the output video, with `--no-ffmpeg-merge` .
* Stream mode now also supports video processing.
* In batch and stream mode, customer can choose to process images and/or videos with `-t image` and `-t video`. Default is `-t image` without `-t video,` `--video` is alias for `-t video` without `-t image`.
* FIXED: Wrong color channels causes degraded detection performance.

**Known bugs**:

* <mark style="color:$info;">FIXED in v26.02.1: Overlapping instance segmentation masks mix up the colours of the instances.</mark>
* <mark style="color:$info;">FIXED in v26.02.2:</mark> <mark style="color:$info;">In REST API mode, the DELETE request does not delete the metadata JSON of video files in the output folder within the container.</mark> &#x20;

## 2025 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

### Version 25.11.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Released on November 13th, 2025.

* Update of the Docker base image to Ubuntu 24.04 with Python 3.12

### Version 25.10.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Released on October 1st, 2025.

* Bugfix: Gradual blur now works for license plate blurring.
* Bugfix: `--face-anonymization-gradient-stop` now works as intended.
* Bugfix: Issue with datetime timezones in telemetry.
* Bugfix: `license` command works now as intended.
* Bugfix: In video processing, metadata JSON are now written in 500-frame chunks (500 frames per JSON file).

### Version 25.09.1 (preview) <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Released on September 8th, 2025.

**Attention**: This release is based on major rewrite of our code base. It contains some breaking changes.

#### Hardware support&#x20;

* Support for NVIDIA Hooper and Blackwell architecture.
* Faster performance.
* Improve hardware utilisation (lower GPU memory consumption).

#### ML model updates

* `segmentatio-omnium` is now the default model! \
  The default thresholds for face and license-plate are now 0.2.
* :exclamation: Model `object-detection-v2`  is **removed**, use `segmentatio-omnium` instead.
* CLI parameters `--model-segmentation` , `----model-license-plate` , `--model-keypoint` are deprecated.
* Model retrained.

#### Other improvements

* Improve anonymisation with smoother gradual blur.
* Instance segmentation masks (`--save-mask instance`) have more distinct colors.
* Stream mode (`--stream`) now also supports video processing!
* Fix bugs in TCP modes:
  * `--save-mask binary` now returns binary mask instead of  blurred image.
  * `--debug` `--score` now also works in TCP modes.
* :exclamation: In Container API synchronous mode, a `POST` request is rejected, if an anonymisation process is ongoing.

#### Obsolete parameters

Following parameters can be removed without any change in behaviour.

* CLI parameter `--metrics` is removed.
* CLI parameter `--non-parallel` is removed.

#### :exclamation:Known bugs

* <mark style="color:$info;">\[FIXED in v25.10.1] Gradual blur does not work for license plate anonymisation.</mark>
* <mark style="color:$info;">\[FIXED in v25.10.1] CLI parameter</mark> <mark style="color:$info;"></mark><mark style="color:$info;">`--face-anonymization-gradient-stop`</mark> <mark style="color:$info;"></mark><mark style="color:$info;">does not work.</mark>
* <mark style="color:$info;">\[FIXED in v25.10.1]</mark> <mark style="color:$info;"></mark><mark style="color:$info;">`license`</mark> <mark style="color:$info;"></mark><mark style="color:$info;">subcommand does not work.</mark>
* <mark style="color:$info;">\[FIXED in v25.10.1] Issue with datetime timezone in telemetry.</mark>&#x20;
* <mark style="color:$info;">\[FIXED in v26.02.2]</mark> <mark style="color:$info;">In Container API mode, when processing a defect video,</mark>  <mark style="color:$info;"></mark><mark style="color:$info;">`status`</mark>  <mark style="color:$info;"></mark><mark style="color:$info;">is "done" instead of "failed".</mark>
* <mark style="color:$info;">\[FIXED in v25.10.1] In video processing, metadata are not written in 500-frame chunks (500 frames per file) as it should be.</mark>
* <mark style="color:$info;">\[FIXED in v26.01.1] Wrong color channels causes degraded detection performance.</mark>

### Version 25.05.1 (hotfix) <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on May 19th, 2025

* Fix issue in Container API caused by processing corrupt video files.

### Version 25.04.2 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on April 23, 2025

* **Improved anonymization quality of** `segmentatio-omnium` **model:**\
  Images taken by UAVs / drones.

### Version 25.04.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on April 8, 2025

* **Improved anonymization quality of** `segmentatio-omnium` **model:**\
  High-resolution Mobile Mapping panoramas, Dashcam images (California/USA)

### Version 25.03.1 (hotfix)

Release on March 18, 2025

* Display file name of the video being processed in log
* Fixed: Graceful shutdown in Batch mode (ctrl+c)
* Fixed: Face blur appearance&#x20;
* Fixed: error in video object tracking
* Fixed: Wrong parameter to --tcp-jpeg terminates the container but does not print out any log
* Fixed: Container API displays non-user-friendly error message if -i and -o parameters are not specified
* Fixed: In Container API, `GET /file/1/anonymised` after setting method to "detect" returns non-anonymized image

### Version 25.01.2 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on February 18, 2025

* Fixed issue with faulty tiling parameter
* Fixed tiling errors not being displayed
* Fixed `GET /v1/status` not responding if a non-sync request is running.
* Removed `--prioritize-nvenc` flag
* Ensure graceful shutdown on Ctrl-C in API mode.
* Removed `--object-tracking` flag from help (by default on, disable with `--no-object-tracking`).

### Version 25.01.1 (hotfix) <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on January 13, 2025

* Fixed bug that made the Container stop unexpectedly and without logging any errors.
* Minor improvements of `object-detection-v2` model.

## 2024 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

### Version 24.10.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on October 31, 2024

* **Improved overall detection quality of** `object-detection-v2` **model (with a focus on imagery from the US)**\
  Generally higher recall and precision over all tested use cases and scenarios.
* **Reduced memory usage**\
  Lower memory consumption for large images and panoramas due to optimized storage of segmentation masks.
* **Preserve IPTC metadata in anonymized images**\
  Copy over essential IPTC metadata from original to anonymized JPEG images.

### Version 24.08.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on August 20, 2024

* Improved logging, updated logging format and hid irrelevant log messages:
  * Logs are colored in STDOUT depending on the log level.
  * Log level can be set via `--log-level` parameter.
  * In REST API mode, the HTTP requests are not logged anymore.
* `object-detection-v2` model now also supports segmentation masks with `--save-mask`.
* <mark style="color:$info;">Known bug: The combination of</mark> <mark style="color:$info;"></mark><mark style="color:$info;">`--model object-detection-v2`</mark> <mark style="color:$info;"></mark><mark style="color:$info;">and</mark> <mark style="color:$info;"></mark><mark style="color:$info;">`--cpu-mode`</mark> <mark style="color:$info;"></mark><mark style="color:$info;">causes error.</mark>

### Version 24.05.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on June 17, 2024

* Remove "ghost" detections when object tracking is enabled for video anonymisation.
* Fix bug with NVIDIA driver caused by conflicting library versions.
* Software maintenance (refactoring) and dependency upgrades

### Version 24.04.1 (hotfix) <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on April 7, 2024&#x20;

* Fix bug: 'color' output in metadata JSON is null instead of RGB triplet for `--save-mask all` and `--save-mask instance`.
* Fix bug: In API mode, response to `POST /v1/file` does not contain `anonymised_url`
* When API mode starts without `--save-mask` parameter, retrieving segmentation masks with `GET /file/{id}/binary-mask` or `GET /file/{id}/instance-mask` now causes 404 instead 500 error.
* Fix bug: In API mode, `DELETE /task/{id}` does not remove video metadata json files.

### Version 24.03.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on March 28, 2024&#x20;

* Improved `object-detection-v2` model.
* Fixed bug with REST API query parameter naming: Dashes `-` are now interchangeable with underscores `_` in parameter names.
* XMP metadata are now cloned into output images: [#will-image-exif-and-xmp-metadata-be-carried-over-to-the-anonymized-image](https://doc.celantur.com/usage/batch-and-stream-mode#will-image-exif-and-xmp-metadata-be-carried-over-to-the-anonymized-image "mention")
* The Docker container now starts by default as user with uid `1000` and gid `1000`.

### Version 24.02.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on Febuary 20, 2024

* Improve UX of Container API asynchronous processing.

### Version 24.01.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on Jan 18, 2024&#x20;

* API mode now supports **asynchronous processing**.
* API mode now supports **webhooks**.
* Improved `object-detection-v2` model.
* In API mode, the `-i <input>` and `-o <output>` folders are used to temporary store the input and output data.

{% hint style="info" %}
If you are using API mode, **remove** the following two lines from`celantur.sh` to avoid an PermissionError:

`-v "${PROCESS_DIR}/input":"${DOCKER_HOME}/input"`

`-v "${PROCESS_DIR}/output":"${DOCKER_HOME}/output"`
{% endhint %}

* In API mode, the keys in JSON response to `POST /v1/file`  now contain underscore instead of dashes.

  ```json
  {
    "id": "string",
    "content_type": "before: content-type",
    "anonymized_video": "before: anonymized-video",
    "metadata_url": "before: metadata-url"
  }
  ```

  &#x20;
* Misc. bugfixes

## 2023

### Version 23.12.1 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on Dec 7, 2023&#x20;

* Improved `object-detection-v2` model.
* Bugfix: Fetching video metadata as JSON returned no detections in certain cases.

*Known issue (Fixed in version 24.02.1): In REST API mode and using the* `object-detection-v2` *model, when submitting a video followed by an image an error occurs (500 HTTP response, "*&#x45;rror in ASGI Framework" in Container log)*. We'll address this issue with a hotfix or in the upcoming release.*&#x20;

### Version 23.11.2 <a href="#draft-version-23.11.2" id="draft-version-23.11.2"></a>

Release on Nov 14, 2023 (Bugfixes)

* Bugfix: Container API CLI parameters are not overriden by POST query parameters.
* Bugfix: Video processing causes crash when video file has no metadata.

### Version 23.11.1 <a href="#draft-version-23.11.1" id="draft-version-23.11.1"></a>

Release on Nov 7, 2023

* General license plate detection improvements (reduction of false positives).
* The `object-detection-v2` model has been retrained on UK license plates.

### Version 23.10.2 <a href="#draft-version-23.10.2" id="draft-version-23.10.2"></a>

Release on Oct 3, 2023

* Integrated object tracking for video processing.
* Keep [video metadata](https://doc.celantur.com/container/usage/batch-and-stream-mode#is-my-video-metadata-carried-over-to-the-anonymized-video) (stream/track and container metadata) when anonymizing videos
* Create [video metadata](https://doc.celantur.com/usage/segmentation-masks-and-metadata#video-metadata) (frame-wise) as JSON of anonymized videos

### Version 23.09.2 <a href="#draft-version-23.09.1" id="draft-version-23.09.1"></a>

Released on October 03, 2023

* In Container REST API, the type of the attribute `id` in the response of `POST /file` changes from number to string.&#x20;

### Version 23.09.1 <a href="#draft-version-23.09.1" id="draft-version-23.09.1"></a>

Released on Sep 12, 2023

* Minor internal changes.

### Version 23.08.2 <a href="#draft-version-23.08.2" id="draft-version-23.08.2"></a>

Released on Sep 5, 2023

* Integrate new video processing model.
* Refactoring of OpenCV dependencies for video processing.

### Version 23.08.1 <a href="#draft-version-23.08.1" id="draft-version-23.08.1"></a>

Released on Aug 23, 2023

* Minor internal changes

### Version 23.07.3 <a href="#draft-version-23.07.3" id="draft-version-23.07.3"></a>

Released on Aug 14, 2023

* Bugfix: Licensing did not work in some cases.

### Version 23.07.2 <a href="#draft-version-23.07.2" id="draft-version-23.07.2"></a>

Released on Jul 22, 2023

* Introduce functionality to limit container usage via license files.
* Misc. bug fixes.

### Version 23.07.1 <a href="#draft-version-23.07.1" id="draft-version-23.07.1"></a>

Released with 23.07.2

* Refactor memory allocation for multiprocessing.
* Refactor image data handling.
* Known issue: `--model object-detection` does not work with shared memory (`--use-shared-memory`)

### Version 23.06.2

Released: 2023-07-13

* Fixed: Wrong bounding boxes in metrics output.
* Fixed: Logging didn't write to main log.
* Fixed: Video processing didn't accept `--debug` `--score` parameters.
* Fixed: `--stream --overwrite` crashes when the output directory contained the file that is named the same as input file.
* Fixed: Misc. vulnerabilities in Celantur Container.
* Fixed: Video processing cannot write out video files if output folder does not exit.
* Minor bugfixes and quality of life improvements.

### Version 23.04.3

Released: 2023-04-30

* Minor bugfixes and quality of life improvements.

### Version 23.03.2

Released: 2023-03-31

* Video anonymization supports now existing inference threshold parameters, e.g. `--<object>-threshold`.
* Fixed: `--debug --score` parameters did not work properly.
* Fixed: blur function caused process to stop in certain edge cases.

### Version 23.02.1

Released: 2023-02-28

* Minor bugfixes and quality of life improvements.

### Version 23.01.2

Released: 2023-01-31

* Fixed: Segmentation and binary masks were blank.

## 2022

### Version 22.12.1

Released: 2022-12-31

* **Customizable Blurring**\
  The blurring effect can now be customized by setting gradient start and stop value, as well as the kernel size for specific objects. [Read how it works.](https://www.celantur.com/doc/beautify-blur/).
* Code refactoring and maintenance (anonymiser code).
* Fixed issue with garbage collection.

### Version 22.11.2

Released: 2022-11-30

* **Get segmentation masks in NumPy TCP mode**\
  By setting the `--save-mask` parameter, segmentation masks can now be returned by the Container (default are numpy arrays).
* Fixed: Celantur Container new logs to STDOUT instead of STDERR.

### Version 22.10.2

Released: 2022-11-03

* Code refactoring to improve performance for video processing (on GPU).
* Fixed: Binary masks contained vehicles and persons, even though only license plates were specified.

### Version 22.09.2

Released: 2022-09-30

* **File-wise detection thresholds**\
  Set individual detection thresholds per file by specifying `face-threshold`, `person-threshold`, `vehicle-threshold` and `license-plate-threshold` when using the [Container API](https://www.celantur.com/doc/container-api/) or [Batch Mode](https://www.celantur.com/doc/on-premise/usage/).
* Container API health endpoint (/status) is now non-blocking.
* Fixed: Automatic tiling.
* Fixed: When anonymising person and license-plate, persons are not anonymised.
* Updated celantur-core dependencies.

### Version 22.08.1

Released: 2022-08-31

* Code refactoring to improve software maintenance.

### Version 22.07.2

Released: 2022-07-31

* **PNG support for Container API**\
  PNG images can be ingested with the Container API.
* Several bug fixes and quality of life improvements.

### Version 22.06.8

Released: 2022-06-30

* **REST API**\
  You can now send and receive images from the container via a REST API. This makes integration into your data pipelines even easier and more convenient: [Celantur Container API Documentation](https://www.celantur.com/doc/container-api/)
* **Generate segmentation masks**\
  Create binary or instance segmentation masks for processed images. These mask files can be used for further image analysis and faster processing: [Generate segmentation masks with the Celantur container](https://www.celantur.com/doc/segmentation-masks/)
* **Up to 60% faster processing by generating only metadata and segmentation masks**\
  By selecting the new “detect” method (instead of e.g. “blur”), the container is running only the inference on images, skipping the anonymization. This results in faster processing times when only metadata or segmentation masks are required.
* **Size of Docker image reduced by 1 GiB**
* Several bug fixes and quality of life improvements.