For the complete documentation index, see llms.txt. This page is also available as Markdown.

Comparison

Celantur vs. EgoBlur Gen 2

EgoBlur Gen 2 is an open-source AI model from Meta. Comparison conducted during July 2026.

Setup

The EgoBlur repo got a little bit refactored which made the setup easier than Gen 1 and it now works without any error.

Usage

With EgoBlur2 you can still only anonymize faces and license plates and no whole vehicles and persons. The clear focus of these models are with the Meta Glasses, since now you can set a specific set of cameras during inference for some finetuned confidence thresholds.

The demo script still only processes a single image at a time and custom logic is also required to read folders, tile large images, save metadata, etc.

Results

Celantur Container 26.04.1 vs EgoBlur Gen 2

Compared to the Celantur Container, EgoBlur Gen 2 shows a similar or slightly lower Recall for license-plates and a significantly lower one for heads. With a few exceptions, the Precision of the Celantur Container for heads is higher across all reference datasets. With a few exceptions, the Precision is also higher for license-plates.

The End-to-End processing times, including the whole start up, model loading, etc., are similar, although the container seems to be a little bit slower. This overhang becomes less the more images are processed. The processing times per image, including inference and anonymization, are significantly faster with the Celantur Container, especially for large images where tiling is necessary. In such cases the container is up to 3x faster than EgoBlur Gen 2. It has to be noted that EgoBlur Gen 2 got ~3x faster compared to the 1st Gen. version.

Even though EgoBlur Gen 2 needs to load two roughly 430MB models into VRAM, the VRAM usage during inference is a little bit lower, especially for large panorama images. The higher VRAM usage from the Container is probably caused by the additional segmentation masks that the Celantur model generates.

Test system specification: RTX 4090 24 GB, Intel Core i7-13700, 64 GB RAM (2x32)

Dataset

VRAM Consumption Celantur Container

VRAM Consumption EgoBlur Gen 2

Total Processing Time Celantur Container

Total Processing Time EgoBlur Gen 2

Processing Time per Image Celantur Container [sec.]

Processing Time per Image EgoBlur Gen 2[sec.]

Dashcam

100 images x 1920x1080

1546

1796

00: 00 : 13

00: 00 : 12

0.043 ± 0.115

0.087 ± 0.112

Dashcam

50 images x 3840x2160

3450

1730

00 : 00 : 14

00 : 00 : 10

0.096 ± 0.162

0.128 ± 0.158

Indoor Mapping

100 images x 3648x5472

1808

2680

00 : 00 : 37

00 : 01 : 05

0.242 ± 0.154

0.658 ± 0.199

Drone/UAV

50 images x 4032x3024

1872

2168

00 : 00 : 24

00 : 00 : 42

0.215 ± 0.164

0.684 ± 0.264

Mobile Mapping

50 panos x 8000x4000

2110

2022

00 : 00 : 16

00 : 00 : 16

0.213 ± 0.274

0.519 ± 0.405

Mobile Mapping

50 panos x 12288x6144

2838

2494

00 : 00 : 46

00 : 00 : 59

0.287 ± 0.163

0.789 ± 0.275

Mobile Mapping

50 panos x 13504x6752

4226

2590

00 : 00 : 54

00 : 01 : 04

0.314 ± 0.161

0.846 ± 0.217

Conclusion

EgoBlur Gen 2, by default, is very limited in its capabilities and requires some coding proficiency to get the desired results. Some drawbacks include:

  • Limited functionality The default sample script only processes a single image at a time. The only output of the sample script is the anonymized version of the input image. Basic functionality like image tiling for larger images, scanning folders for multiple images, saving metadata, customizing the appearence of the anonymization is all missing and must be self implemented.

  • Lackluster Results While the detection capabilities for license-plates are comparable to the Celantur Container, Heads are repeatedly not detected. Furthermore, EgoBlur2 produces significantly more false-positive detections than Celantur Container, reducing the overall quality of the end result.

While EgoBlur Gen 2 is considerable faster than the 1st generation model, it is still plagued by many of the same problems and it remains slower than Celantur Container.

Celantur vs. EgoBlur Gen 1

EgoBlur is an open-source AI model from Meta. Comparison conducted during October 2025.

Setup

The setup turned out to be surprisingly difficult. In the official repo they provide an environment.yaml file to create a new conda environment. This didn’t work at all on our servers. Constant problems with the specified versions of OpenCV, Numpy and torch. Only solution was to create an environment from scratch with Python's venv.

Usage

With EgoBlur only faces and license plates can be anonymized, no whole vehicles and persons. Furthermore you can only process 1 image at a time with the provided sample script and need custom logic to: read folders, tile large images, save metadata, etc.

Results

Compared to the Celantur Container, EgoBlur shows a similar Recall for license-plates but a lower one for Heads. The Precision of the Celantur Container across all tested datasets is significantly better. The End-to-End processing times, including the whole start up, model loading, etc., are similar, although the container seems to be a little bit slower. This is to be expected I guess since we have a lot more features and logic to handle.

The processing times per image, including inference and anonymization, are significantly faster with the Celantur Container. The VRAM usage for both solutions is quite similar.

Test system specification: RTX 4090 24 GB, Intel Core i7-13700, 64 GB RAM (2x32)

Dataset
VRAM Consumption Celantur Container
VRAM Consumption EgoBlur
Celantur consumes x% less VRAM
Total Processing Time [HH : MM : SS]
Total Processing Time [HH : MM : SS]
Celantur x% faster
Processing Time per Image [sec.]
Processing Time per Image [sec.]

Dashcam 100 images x 1920x1080

1.2

3.8

68%

00: 00 : 07

00: 00 : 20

65%

0.048 ± 0.132

0.164 ± 0.198

Dashcam 50 images x 3840x2160

2.7

11.2

76%

00 : 00 : 10

00 : 00 : 35

71%

0.113 ± 0.187

0.610 ± 0.250

Indoor Mapping 100 images x 3648x5472

1.4

7.9

82%

00 : 00 : 36

00 : 03 : 34

83%

0.280 ± 0.146

1.877 ± 0.279

Drone/UAV 50 images x 4032x3024

1.1

9.3

88%

00 : 00 : 21

00 : 01 : 31

77%

0.283 ± 0.182

1.544 ± 0.492

Mobile Mapping 50 panos x 8000x4000

1.4

7.7

82%

00 : 00 : 35

00 : 01 : 42

66%

0.534 ± 0.202

1.513 ± 0.382

Mobile Mapping 50 panos x 12288x6144

1.6

12.1

87%

00 : 00 : 45

00 : 03 : 19

77%

0.733 ± 0.210

3.049 ± 0.347

Mobile Mapping 50 panos x 13504x6752

2.0

12.8

84%

00 : 00 : 54

00 : 03 : 41

76%

0.924 ± 0.242

3.326 ± 0.358

Conclusion

EgoBlur, by default, is very limited in its capabilities and requires software engineering proficiency to get the desired results. Some drawbacks include:

  • Limited functionality The default sample script only processes a single image at a time. The only output of the sample script is the anonymized version of the input image. Basic functionality like image tiling for larger images, scanning folders for multiple images, saving metadata, customizing the appearance of the anonymization is all missing and must be self implemented.

  • Cumbersome Setup The setup of the environment to run EgoBlur also expects the user to know what a conda environment is and how to set it up. The provided environment.yaml file is not fail-proof however and to fix any issues one must again have the knowledge to debug the environment and code.

  • Lacklaster Results While the detection capabilities for license-plates are comparable to the Celantur Container, Heads are repeatedly not detected. Furthermore, EgoBlur produces significantly more false-positive detections than Celantur Container, reducing the overall quality of the end result.

Last updated