An extendable tool for automatic annotation of image data by a combination of deep neural networks.
The primary objective of this annotator is to prioritize the accuracy and quality of predictions over speed. The autoannotator
has been specifically designed to surpass the precision offered by most publicly available tools. It leverages ensembles of deep neural models to ensure the utmost quality in its predictions. It is important to note that neural networks trained on clean datasets tend to yield superior results compared to those trained on larger but noisier datasets.
- Face and landmarks detection
- Face alignment via keypoints
- Face descriptor extraction
- Clusterization
Task | Hardware | Time, s |
---|---|---|
Face detection + landmarks + extraction | Xeon e5 2678 v3 | ~1 |
pip install autoannotator
Check out our demo face recognition pipeline at: examples/face_recognition_example.py
git clone https://github.com/CatsWhoTrain/autoannotator_client
cd autoannotator_client
docker compose up
The webinterface could be found locally at: http://localhost:8080/
We have asked engineers in the field of video analytics whether they are interested in such a library. Their responses were:
- IREX: would use this library and contribute to it.
- NapoleonIT: would use this library and contribute to it.
- ITMO.Lens: would use this library.
What are the reasons for choosing this data labeling tool over the alternative of employing human annotators?
Long time ago Andrej Karpathy observed that his accuracy was only 94% when he tried to label just 400 images of the CIFAR-10 dataset while SOTA Efficient adaptive ensembling for image classification (August 29, 2023) achieves >99.6% accuracy. When expert labelers had to choose from ~100 labels while annotating ImageNet, their error rate increased to 13-15%.
Andrej's error rate was determined to be 5.1%, and he initially invested approximately one minute in labeling a single image. Conversely, utilizing Florence or never models for the same task can deliver a top-5 error rate of less than 1%.
A certain undisclosed company, bound by a non-disclosure agreement (NDA), has utilized a technique wherein face images captured under challenging environmental conditions are pre-processed. This procedure involves the application of both a facial recognition network and DBSCAN algorithm to divide the images into distinct individuals. Subsequently, human annotators undertook a validation process to verify the accuracy of the pre-processed data. The work conducted by the annotators was inspected by their team leader. Ultimately, it was determined by an ML engineer that 1.4% of the clustered face images were mislabeled.
Current repository takes ideas and some code form the following projects:
All individuals depicted in the "assets" images have provided explicit consent to utilize their photos for the purpose of showcasing the current library's work. Kindly refrain from utilizing these images within your individual projects.