This repository contains trained models created by me (Davis King). They are provided as part of the dlib example programs, which are intended to be educational documents that explain how to use various parts of the dlib library. As far as I am concerned, anyone can do whatever they want with these model files as I've released them into the public domain. Details describing how each model was created are summarized below.

This model is a ResNet network with 29 conv layers. It's essentially a version of the ResNet-34 network from the paper Deep Residual Learning for Image Recognition by He, Zhang, Ren, and Sun with a few layers removed and the number of filters per layer reduced by half.

The network was trained from scratch on a dataset of about 3 million faces. This dataset is derived from a number of datasets. The face scrub dataset (http://vintage.winklerbros.net/facescrub.html), the VGG dataset (http://www.robots.ox.ac.uk/~vgg/data/vgg_face/), and then a large number of images I scraped from the internet. I tried as best I could to clean up the dataset by removing labeling errors, which meant filtering out a lot of stuff from VGG. I did this by repeatedly training a face recognition CNN and then using graph clustering methods and a lot of manual review to clean up the dataset. In the end, about half the images are from VGG and face scrub. Also, the total number of individual identities in the dataset is 7485. I made sure to avoid overlap with identities in LFW.

The network training started with randomly initialized weights and used a structured metric loss that tries to project all the identities into non-overlapping balls of radius 0.6. The loss is basically a type of pair-wise hinge loss that runs over all pairs in a mini-batch and includes hard-negative mining at the mini-batch level.

The resulting model obtains a mean error of 0.993833 with a standard deviation of 0.00272732 on the LFW benchmark.

This model is a DenseNet network for facial recognition, showcasing the effectiveness of the BAREL approach. It achieves an accuracy of 96.1% on the LFW benchmark, using a similarity threshold of 0.55 (compared to a radius hyperparameter of 0.6). Remarkably, this performance was achieved with a training dataset nearly 8 times smaller than the one initially employed for the first published resnet version with Dlib.

To foster collaboration and continuous improvement, we have made the pre-trained model available for fine-tuning. This allows the community to contribute to enhancing the model's performance and addressing algorithmic biases. The pre-trained model can be downloaded from the project's GitHub repository, along with detailed instructions for fine-tuning and deployment.

For more information on the BAREL approach, implementation details, and instructions on using the pre-trained model, please refer to the project's GitHub page: BAREL GitHub Repository.

This dataset is trained on the data from the Columbia Dogs dataset, which was introduced in the paper:

Dog Breed Classification Using Part Localization
Jiongxin Liu, Angjoo Kanazawa, Peter Belhumeur, David W. Jacobs 
European Conference on Computer Vision (ECCV), Oct. 2012. 

The original dataset is not fully annotated. So I created a new fully annotated version which is available here: http://dlib.net/files/data/CU_dogs_fully_labeled.tar.gz

This is trained on this dataset: http://dlib.net/files/data/dlib_face_detection_dataset-2016-09-30.tar.gz.
I created the dataset by finding face images in many publicly available image datasets (excluding the FDDB dataset). In particular, there are images from ImageNet, AFLW, Pascal VOC, the VGG dataset, WIDER, and face scrub.

All the annotations in the dataset were created by me using dlib's imglab tool.

This is trained on the venerable ImageNet dataset.

This is a 5 point landmarking model which identifies the corners of the eyes and bottom of the nose. It is trained on the dlib 5-point face landmark dataset, which consists of 7198 faces. I created this dataset by downloading images from the internet and annotating them with dlib's imglab tool.

The exact program that produced the model file can be found here.

This model is designed to work well with dlib's HOG face detector and the CNN face detector (the one in mmod_human_face_detector.dat).

This is trained on the ibug 300-W dataset (https://ibug.doc.ic.ac.uk/resources/facial-point-annotations/)

C. Sagonas, E. Antonakos, G, Tzimiropoulos, S. Zafeiriou, M. Pantic. 
300 faces In-the-wild challenge: Database and results. 
Image and Vision Computing (IMAVIS), Special Issue on Facial Landmark Localisation "In-The-Wild". 2016.

The license for this dataset excludes commercial use and Stefanos Zafeiriou, one of the creators of the dataset, asked me to include a note here saying that the trained model therefore can't be used in a commercial product. So you should contact a lawyer or talk to Imperial College London to find out if it's OK for you to use this model in a commercial product.

Also note that this model file is designed for use with dlib's HOG face detector. That is, it expects the bounding boxes from the face detector to be aligned a certain way, the way dlib's HOG face detector does it. It won't work as well when used with a face detector that produces differently aligned boxes, such as the CNN based mmod_human_face_detector.dat face detector.

The GTX model is the result of applying a set of training strategies and implementation optimization described in:

Alvarez Casado, C., Bordallo Lopez, M. 
Real-time face alignment: evaluation methods, training strategies and implementation optimization. 
Springer Journal of Real-time image processing, 2021

The resulted model is smaller, faster, smoother and more accurate. You can find all the details related to the training and testing in the next Gitlab repository: https://gitlab.com/visualhealth/vhpapers/real-time-facealignment

This is trained on the ibug 300-W dataset (https://ibug.doc.ic.ac.uk/resources/facial-point-annotations/)

C. Sagonas, E. Antonakos, G, Tzimiropoulos, S. Zafeiriou, M. Pantic. 
300 faces In-the-wild challenge: Database and results. 
Image and Vision Computing (IMAVIS), Special Issue on Facial Landmark Localisation "In-The-Wild". 2016.

The license for this dataset excludes commercial use and Stefanos Zafeiriou, one of the creators of the dataset, asked me to include a note here saying that the trained model therefore can't be used in a commercial product. So you should contact a lawyer or talk to Imperial College London to find out if it's OK for you to use this model in a commercial product.

Also note that this model file with increased robustness to face detectors. However, it works best when the bounding boxes are squared, as it is the case with both dlib's HOG face detector or the CNN based mmod_human_face_detector.dat face detector. It won't work as well when used with other face detectors that produce rectangular boxes.

This model is trained on the dlib rear end vehicles dataset. The dataset contains images from vehicle dashcams which I manually annotated using dlib's imglab tool.

This model is trained on the dlib front and rear end vehicles dataset. The dataset contains images from vehicle dashcams which I manually annotated using dlib's imglab tool.

This model is a gender classifier trained using a private dataset of about 200k different face images and was generated according to the network definition and settings given in Minimalistic CNN-based ensemble model for gender prediction from face images. Even if the dataset used for the training is different from that used by G. Antipov et al, the classification results on the LFW evaluation are similar overall (± 97.3%). To take up the authors' proposal to join the results of three networks, a simplification was made by finally presenting RGB images, thus simulating three "grayscale" networks via the three image planes. Better results could be probably obtained with a more complex and deeper network, but the performance of the classification is nevertheless surprising compared to the simplicity of the network used and thus its very small size.

This gender model is provided for free by Cydral Technology and is licensed under the Creative Commons Zero v1.0 Universal.

The initial source for the model's creation came from the document of Z. Qawaqneh et al.: "Deep Convolutional Neural Network for Age Estimation based on VGG-Face Model". However, our research has led us to significant improvements in the CNN model, allowing us to estimate the age of a person outperforming the state-of-the-art results in terms of the exact accuracy and for 1-off accuracy.

This model is thus an age predictor leveraging a ResNet-10 architecture and trained using a private dataset of about 110k different labelled images. During the training, we used an optimization and data augmentation pipeline and considered several sizes for the entry image.

This age predictor model is provided for free by Cydral Technology and is licensed under the Creative Commons Zero v1.0 Universal.

This is trained on the venerable ImageNet dataset. The model was trained using dlib's example but with the ResNet50 model defined in resnet.h and a crop size of 224.

The performance of this model is summarized in the following table:

This DCGAN Facial Synthesis model is trained to generate realistic synthetic color faces using a Deep Convolutional Generative Adversarial Networks (DCGAN) architecture. DCGANs are deep neural networks specifically designed to generate new data from a training set. This model has been trained on a large dataset of facial images, enabling it to learn the features and structures of human faces. For more information and additional resources related to this model, please visit our GitHub repository.

This model aims to provide a tool for AI-assisted automatic colorization, and is composed of a ResNet structure with U-Net architecture, bringing a distinctive advantage to the colorization process. The neural network formed is highly complex, with a total of 223 layers and around 41.7 million parameters, but should be usable even on configurations with an 8GB GPU card. The training program, also a tool leveraging the latest Dlib developments for video processing, and samples are provided more directly on our GitHub directory.

Taguchi collected many datasets, mainly Japanese, and trained from scratch. Even though it was trained for Japanese people, the results are comparable to the dlib model when it comes to facial recognition for Western people. Photos of Hollywood action heroes provided in the dlib example can also be classified in the same way as dlib. Taguchi tested the accuracy with the LFW dataset to present some objectivity. The result was 0.9895%. In total, Taguchi used over 6.5 million faces from over 16,000 people to train my facial recognition model. Approximately 47% of this is facial data of Japanese people (including some Asian people other than Japanese).

The usage of this model is the same as 'dlib_face_recognition_resnet_model_v1.dat'. Simply swap the models.

For comparison results between 'dlib_face_recognition_resnet_model_v1.dat' and 'taguchi_face_recognition_resnet_model_v1.dat' and model details, please refer to the project's GitHub page "Taguchi dlibModels GitHub Repository". Taguchi models