Vehicle and lane detection
Currently at Keeptruckin, I am working on vehicle and lane detection to build an understanding of road and objects around the ego-vehicle. Specifically, I am working on building a unified deep architecture for vehicle and lane detection employing cost-sensitive multi-task learning approach.
Oriented object detection
Automatic detection of firearms is important for enhancing security and safety of people, however, it is a challenging task owing to the wide variations in shape, size and appearance of firearms. Moreover, the existing object detectors process rectangular areas and a thin and long rifle may actually cover only a small percentage of that area and the rest may contain irrelevant details suppressing the required object signatures. To handle these challenges we propose an Orientation Aware Object Detector. The proposed detector has achieved improved firearm detection and localization performance compared to the state-of-the art object detectors. You can read more and view some qualitative results on Project page.
Publication:
"Orientation Aware Object Detection with Application to Firearms", under review at IEEE TIP (Github Link)
Automatic detection of firearms is important for enhancing security and safety of people, however, it is a challenging task owing to the wide variations in shape, size and appearance of firearms. Moreover, the existing object detectors process rectangular areas and a thin and long rifle may actually cover only a small percentage of that area and the rest may contain irrelevant details suppressing the required object signatures. To handle these challenges we propose an Orientation Aware Object Detector. The proposed detector has achieved improved firearm detection and localization performance compared to the state-of-the art object detectors. You can read more and view some qualitative results on Project page.
Publication:
"Orientation Aware Object Detection with Application to Firearms", under review at IEEE TIP (Github Link)
Affective Image Transformation
I have also worked on transferring the evoked emotion of an image by manipulating its low level features while keeping in consideration the content of both source and target images. The proposed approach allow much more control of desired emotion by providing user ability to pick any discrete distribution of the emotions. The idea behind the approach was to show that where low-level-features assign the manipulative power over the image, content of image also constraints how much the image can be changed. You can read more about the proposed methodology and view the qualitative results on project page.
Publication:
"Automatic Image transformation for Inducing Affect", British Machine Vision Conference, 2017
I have also worked on transferring the evoked emotion of an image by manipulating its low level features while keeping in consideration the content of both source and target images. The proposed approach allow much more control of desired emotion by providing user ability to pick any discrete distribution of the emotions. The idea behind the approach was to show that where low-level-features assign the manipulative power over the image, content of image also constraints how much the image can be changed. You can read more about the proposed methodology and view the qualitative results on project page.
Publication:
"Automatic Image transformation for Inducing Affect", British Machine Vision Conference, 2017
Affective Image Analysis
My Master's thesis was on affective understanding of images. Unlike existing work that only uses low level visual features, I focused on use of interpret-able high level concepts like objects, places and relationship among them to model human emotions. This is done by utilizing high level semantic information produced by deep learning based object and place detectors . You can read more and view some qualitative results on project page.
Publication:
"High-Level Concepts for Affective Understanding of Images", IEEE Winter Conference on Applications of Computer Vision, 2017
My Master's thesis was on affective understanding of images. Unlike existing work that only uses low level visual features, I focused on use of interpret-able high level concepts like objects, places and relationship among them to model human emotions. This is done by utilizing high level semantic information produced by deep learning based object and place detectors . You can read more and view some qualitative results on project page.
Publication:
"High-Level Concepts for Affective Understanding of Images", IEEE Winter Conference on Applications of Computer Vision, 2017
Image Colorization Using Deep Convolutional Networks
A colored image of a person’s face contains important information like complexion and hair color of the person, while making a sketch we lose this information and are only left with recognizable features. Our goal in this Deep Learning course project was to replicate the ECCV 2016 paper Convolutional Sketch Inversion that introduced a Deep convolution neural network (DNN) to generate photo-realistic images from face sketches. Doing so, we came across many issues and therefore learn alot. We used Chainer- framework deep learning for the implementation of this network. You can read more and view some qualitative results here.
A colored image of a person’s face contains important information like complexion and hair color of the person, while making a sketch we lose this information and are only left with recognizable features. Our goal in this Deep Learning course project was to replicate the ECCV 2016 paper Convolutional Sketch Inversion that introduced a Deep convolution neural network (DNN) to generate photo-realistic images from face sketches. Doing so, we came across many issues and therefore learn alot. We used Chainer- framework deep learning for the implementation of this network. You can read more and view some qualitative results here.
Speech Recognition
speech recognition and developed automatic speech recognition systems (ASRs) and voice activity detectors for Mobile based Urdu Spoken Dialog systems that provides weather information and location-based services to low literate population of Pakistan. I have worked on different HMM based speech recognition toolkits such as CMU Sphinx, HTK and Julius. I have also worked on the factors responsible for drop in ASR accuracy in field environment such as speaker variation, accent variation, poor voice activity detection, environmental noise.
Publication:
"Accent Classification among Punjabi, Urdu, Pashto, Saraiki and Sindhi Accents of Urdu Language", Conference on Language and Technology, 2014