Yash Khandhediya

Ayesha Gurnani, Kenil Shah, Vandit Gajjar et al.

How can we improve the facial soft-biometric classification with help of the human visual system? This paper explores the use of saliency which is equivalent to the human visual system to classify Age, Gender and Facial Expression soft-biometric for facial images. Using the Deep Multi-level Network (ML-Net) [1] and off-the-shelf face detector [2], we propose our approach - SAF-BAGE, which first detects the face in the test image, increases the Bounding Box (B-Box) margin by 30%, finds the saliency map using ML-Net, with 30% reweighted ratio of saliency map, it multiplies with the input cropped face and extracts the Convolutional Neural Networks (CNN) predictions on the multiplied reweighted salient face. Our CNN uses the model AlexNet [3], which is pre-trained on ImageNet. The proposed approach surpasses the performance of other approaches, increasing the state-of-the-art by approximately 0.8% on the widely-used Adience [28] dataset for Age and Gender classification and by nearly 3% on the recent AffectNet [36] dataset for Facial Expression classification. We hope our simple, reproducible and effective approach will help ease future research in facial soft-biometric classification using saliency.

Vandit Gajjar, Yash Khandhediya, Ayesha Gurnani et al.

The paper presents a technique to improve human detection in still images using deep learning. Our novel method, ViS-HuD, computes visual saliency map from the image. Then the input image is multiplied by the map and product is fed to the Convolutional Neural Network (CNN) which detects humans in the image. A visual saliency map is generated using ML-Net and human detection is carried out using DetectNet. ML-Net is pre-trained on SALICON while, DetectNet is pre-trained on ImageNet database for visual saliency detection and image classification respectively. The CNNs of ViS-HuD were trained on two challenging databases - Penn Fudan and TUD-Brussels Benchmark. Experimental results demonstrate that the proposed method achieves state-of-the-art performance on Penn Fudan Dataset with 91.4% human detection accuracy and it achieves average miss-rate of 53% on the TUDBrussels benchmark.

Yash Khandhediya, Karishma Sav, Vandit Gajjar

Surveillance based on Computer Vision has become a major necessity in current era. Most of the surveillance systems operate on visible light imaging, but performance based on visible light imaging is limited due to some factors like variation in light intensity during the daytime. The matter of concern lies in the need for processing images in low light, such as in the need of nighttime surveillance. In this paper, we have proposed a novel approach for human detection using FLIR(Forward Looking Infrared) camera. As the principle involves sensing based on thermal radiation in the Near IR Region, it is possible to detect Humans from an image captured using a FLIR camera even in low light. The proposed method for human detection involves processing of Thermal images by using HOG (Histogram of Oriented Gradients) feature extraction technique along with some enhancements. The principle of the proposed technique lies in an adaptive background subtraction algorithm, which works in association with the HOG technique. By means of this method, we are able to reduce execution time, precision and some other parameters, which result in improvement of overall accuracy of the human detection system.

Vandit Gajjar, Ayesha Gurnani, Yash Khandhediya

With crimes on the rise all around the world, video surveillance is becoming more important day by day. Due to the lack of human resources to monitor this increasing number of cameras manually new computer vision algorithms to perform lower and higher level tasks are being developed. We have developed a new method incorporating the most acclaimed Histograms of Oriented Gradients the theory of Visual Saliency and the saliency prediction model Deep Multi Level Network to detect human beings in video sequences. Furthermore we implemented the k Means algorithm to cluster the HOG feature vectors of the positively detected windows and determined the path followed by a person in the video. We achieved a detection precision of 83.11% and a recall of 41.27%. We obtained these results 76.866 times faster than classification on normal images.