Detection and Analysis of Objects and Faces
Fraunhofer Institute for Integrated Circuits
Technology
In developing systems for object detection and analysis, we take special care to make them efficient and robust and to ensure they can be generically applied to classes of objects.
Requirements
Today's object detection and analysis systems need to perform well under a wide range of conditions. Among other things, they have to be able to cope with:
- any type of background
- variable object size
- multiple objects
- absence of color information (only gray scale values available)
- changes in lighting conditions
- occlusion
- real-time constraints
Structural features used as landmarks
Using sets of image points known as landmarks, which help capture the textures and contours of faces and other objects, it is possible to achieve a very high degree of robustness to variations in background and lighting conditions. Moreover, the algorithms used to calculate landmarks have been optimized to a very high standard of efficiency.
Machine learning
The use of machine learning algorithms makes it possible to train a method to identify defining structural characteristics so that it can be applied to an entire class of objects. Additionally, these algorithms help increase robustness, as they are able to generalize from a given task. This means that reliable decisions can be made even in unknown situations.
Face detection and in-depth analysis
Within the area of object detection, one of our core R&D themes is face detection and analysis. The results of our work are incorporated into our SHORE™ software solution. For detailed descriptions of SHORE™ and its detection and analysis capabilities please see here .
Our extensive expertise in face detection and analysis comes from many years of experience. Also, we have at our disposal a large pool of data for machine learning purposes, with a dataset of over 10,000 annotated faces available for a process known as training. Based on this data and using a combination of landmark points and learning algorithms, we are able to train models that enable very high detection rates:
| Frontal face detection | 91.5% detection rate, with 10 erroneous detections | CMU+MIT dataset |
| Gender classification | 94.3% accuracy | BioID dataset |
| Age estimation | Mean absolute error in years: 6.85 | FG-NET dataset |
Constant optimization of CPU-intensive components in terms of both algorithms and hardware has enabled real-time image processing. The table below shows the performance and processing speed achieved with various configurations (Intel Core 2 Duo 6420, 1 core used; BioID dataset, 384 x 286 pixels).
| Face detection | X | X | X | X |
| Eye detection | X | X | X | |
| Gender classification | X | X | ||
| Recognition of four facial expressions | X | |||
| Time per frame [ms] | 9.4 | 19.3 | 19.9 | 22.0 |
| Frames per second [fps] | 107.5 | 51.8 | 50.3 | 45.5 |
Set Bookmark