D-LORD: DYSL-AI Database for Low-Resolution Disguised Face Recognition

D-LORD is a large-scale dataset that aims at strengthening the research on low-resolution disguised face detection. It leverages surveillance videos where the subjects may or may not be wearing disguise accessories, such as hats, beards and mustaches, sunglasses, and face masks.

We benchmark the performance of various face detection, face recognition, face restoration, and frame selection algorithms on the proposed D-LORD.

Face Detection

Face Detection is performed on the video frames in-order to extract the region of interest. In this study, we utilize three face detection algorithms:

1. YoloV5Face Github link - https://github.com/deepcam-cn/yolov5-face

   This Detection algorithm took 24.20 ms per frame

2. RetinaFace Github link - https://github.com/elliottzheng/face-detection

   This Detection algorithm took 308.45 ms per frame

3. Img2Pose Github link - https://github.com/vitoralbiero/img2pose

   This Detection algorithm took 78.02 ms per frame

It has been observed that YoloV5Face yields greater Mean IOU(mIOU) for all distances, i.e. 5m (96%), 7m (92%), 10m (89%), and 15m (85%). This performance is better as compared to other Face detection algorithms.

Frame Selection

To select the best frames for face recognition, we divide the surveillance videos into four parts based on the subject's distances from the camera (5m, 7m, 10m, and 15m). From each part, we select three best frames using three frame selection algorithms:

1. DWT Github link - https://ieeexplore.ieee.org/document/7850956

   This algorithm took 7.90 ms per frame

2. FQ Github link - https://github.com/uam-biometrics/FaceQnet

   This algorithm took 50.60 ms per frame

3. SER-FIQ Github link - https://github.com/pterhoer/FaceImageQuality

   This algorithm took 518.00 ms per frame

From our experiments, SER-FIQ proves to be the best frame selection algorithm, but its per frame inference time is 65 times higher than the DWT-based method. SER-FIQ yields 79.35% GAR at 5m and 49.45% GAR at 15m. Further results and bar plots are shown in the paper.

Face Restoration

Face restoration techniques are used to improve the quality of low-resolution facial images:

1. GFP-GAN Github link - https://github.com/TencentARC/GFPGAN

   This algorithm took 11.00 ms per frame

2. PSFR-GAN Github link - https://github.com/chaofengc/PSFRGAN

   This algorithm took 100.00 ms per frame

3. Real-ESRGAN Github link - https://github.com/xinntao/Real-ESRGAN

   This algorithm took 52.20 ms per frame

It is observed that GFP-GAN takes less time for super resolution of the faces.

Face Recognition

The three state-of-the-art Face Recognition algorithms have been taken into consideration for performing the experiments:

1. ArcFace Github link - https://github.com/deepinsight/insightface/tree/master/recognition/arcface_torch

   This algorithm took 4.21 ms per frame

2. LightCNN Github link - https://github.com/AlfredXiangWu/LightCNN

   This algorithm took 67.00 ms per frame

3. ElasticFace Github link - https://github.com/fdbtrs/ElasticFace

   This algorithm took 58.64 ms per frame

It has been observed that ElasticFace performs better at 5m and 7m distances, while ArcFace performs better for distances more than 10m and 15m. The performance of LightCNN is significantly lower than the other recognition models.

The detailed results are discussed in the paper.

Downscaled LR Vs Real LR

We provide some samples of downscaled LR and real LR images in various resolutions. It can be observed that there is a significant difference in the visual quality of real and downscaled images.

Datasheet for D-LORD

Link to Datasheet for D-LORD dataset - link

License Agreement for D-LORD

The requester is expected to get this license agreement signed by the legal authority approved to sign on behalf of the institute. License Agreement