Real-Time Gender Detection in the Wild Using Deep Neural Networks

Abstract: Gender recognition can be used in many applications, such as video surveillance, human-computer interaction and customized advertisement. Current state-of-the-art gender recognition methods are detector-dependent or region-dependent, focusing mostly on facial features (a face detector is typically required). These limitations do not allow an end-to-end training pipeline, and many features used in the detection phase must be re-learned in the classification step. Furthermore, the use of facial features limits the application of such methods in the wild, where the face might not be present. This paper presents a real-time end-to-end gender detector based on deep neural networks. The proposed method detects and recognizes the gender of persons in the wild, meaning in images with a high variability in pose, illumination an occlusions. To train and evaluate the results a new annotation set of Pascal VOC 2007 and CelebA were created. Our experimental results indicate that combining both datasets during training can increase the mAp of our gender detector. We also visually analyze which parts leads our network to make mistakes and the bias introduced by the training data.

Authors: Luis Felipe Zeni and Claudio Jung

Video Demos

Real-time screen capture demo using an webcam and Nvidia GTX 1080.

Proposed method processing a crowded video.

Regions that the network activates during the process of gender detection.

Code and Data

Paper: You can download the full paper [ here ]

If you use our code or our dataset annotations in an academic work cite the following paper:

	author={Luis Zeni and Claudio Rosito Jung}, 
	title={Real-Time Gender Detection in the Wild Using Deep Neural Networks}, 
	keywords={gender detection, deep learning, visualization}, 

Annotations: In this work we adapted the Pascal VOC 2007 and CelebA datasets to the task of gender detection in the wild.

Files and instructions on how to set up both datasets and they respective annotations are found [ here ]

Source Code: This work source code is available on github

Instructions on how to set up and use our code are available on our [ git repo ]

Docker: If you just want to test our models without lots of work configuring libs and stuff we made available an nvidia-docker with everything set up at docker hub.

Link to the docker image [ here ]

It is really easy to reproduce our results using nvidia-docker :)

Instructions to reproduce using docker: [ here ]

Pre-trained models: You can download our pre-trained models to run on darknet or tensorflow.

Darknet: 50% VOC2007 + 50% CelebA: weights - [ weights ] - [ descriptors ]

TensorFlow: 50% VOC2007 + 50% CelebA: - [ weights ] - [ descriptors ]


If you need more information, please feel free to contact me through the following e-mail:

luis [dot] zeni [at] inf [dot] ufrgs [dot] br


We would like to thank the Brazilian funding agencies CAPES and CNPq. As well as NVIDIA Corporation for the donation of the Titan Xp Pascal GPU used for this research.