Paper: D2SC-GAN: Dual Deep-Shallow Channeled Generative Adversarial Network, for Resolving Low-resolution Faces for Recognition in Classroom scenarios
published in
IEEE Transactions on Biometrics, Behavior, and Identity Science
We propose the Indian Classroom Face Dataset (ICFD) which consists of images of 84 subjects captured in both controlled laboratory (for training) and unconstrained classroom settings (for probes), to explore the challenges present in performing face recognition in unconstrained classroom settings for the purpose of classroom attendance and search.
For the gallery set, the face images of the subject were captured in controlled laboratory settings with adequate illumination conditions without any form of real-world perturbations. The subjects were sitted at a distance of around 2 meters from a Nikon D7200 DSLR camera and were asked to move their head to capture images at various pitch and yaw angles.
The probe samples were captured in various classrooms and seminar rooms using the same Nikon D7200 DSLR camera as well as a Samsung S9 mobile-camera, where volunteers were instructed to stare at the dias/screen in front, but also have human movements, to create a realistic classroom scenario.
Videos were captured in two modes:
We also collected probe samples in poor illumination conditions for a subset of 8 subjects, by turning off all lights in the classroom and projecting a blank white slide on the screen in front of the audience, which acted as the sole source of light in the room. Images were captured using both (a) DSLR, and (b) Samsung S9 mobile camera.
Few samples from the probe set of ICFD exhibiting: (a) poor illumination with low contrast/resolution, (b) blur, (c) occlusion and (d) pose variations present in the dataset. The samples in red outline indicate faces captured using a mobile (Samsung S9) camera.
As mentioned previously, the images in probe set of ICFD were collected in 4 settings:
1. Panning a DSLR camera (Nikon D7200), fixed on a tripod at the center of an elevated dais in front of the audience, at a distance of about 3 meters from the nearest volunteer, in well-illuminated conditions.
2. Using a Samsung S9 mobile camera, by moving sideways across the dais in front of the audience, in well-illuminated conditions.
3. Panning a DSLR camera in poor illumination settings by switching off all lights in the classroom and projecting a empty white slide on the projector screen in front of the audience, acting as the sole source of light in the room.
4. Using a Samsung S9 mobile camera in the same settings as in point 3.
All subjects were not available for each of the 4 settings described above. The dataset statistics for each of the separatepartition of the dataset is given in the table below.
Any work made public, whatever the form, based directly or indirectly on any part of the Database should include the following reference:
@article{9050650, author={A. {Bhattacharjee} and S. {Das}}, journal={IEEE Transactions on Biometrics, Behavior, and Identity Science}, title={D2SC-GAN: Dual Deep-Shallow Channeled Generative Adversarial Network, for Resolving Low-Resolution Faces for Recognition in Classroom Scenarios}, year={2020}, volume={2}, number={3}, pages={223-234}, DOI={10.1109/TBIOM.2020.2983524}}