Deep-Learning Features, Graphs and Scene Understanding

Abhijeet Kumar

Abstract

Scene Understanding has been a major aspiration of computer vision from its early days. Its root lies in enabling the computer/robot/machine to understand, interpret and manipulate visual data, in similarity to what an average human eye does in front of a natural/artificial localized location/scene. This ennoblement of the machine have a widespread impact ranging from Surveillance, Aerial Imaging, Autonomous Navigation, Smart Cities and thus scene understanding have remained as an active area of research in the last decade. In the last decade, the scope of problems in the scene understanding community has broadened from Image Annotation, Image Captioning, Image Segmentation to Object Detection, Dense Image Captioning, Instance Segmentation etc. Advanced problems like Autonomous Navigation, Panoptic Segmentation, Video Summarization, Multi-Person Tracking in Crowded Scenes have also surfaced in this arena and are being vigorously attempted. Deep Learning has played a major role in this advancement/development. The performance metrics in some of these tasks have more than tripled in the last decade itself but these tasks remain far from solved. Success originating from deep learning can be attributed to the learned features. In simple words, features learned from a Convolutional Neural Network trained for annotation are in general far more suited for captioning then a non-deep learning method trained for captioning. Taking cue from this particular deep learning trend, we dived into the domain of scene understanding with the focus on utilization of prelearned-features from other similar domains. We focus on two tasks in particular: Automatic (multi-label)Image Annotation and (Road)Intersection Recognition. Automatic image annotation is one of the earliest problems in scene understanding and refers to the task of assigning (multiple) labels to an image based on its content. Whereas intersection recognition is the outcome of the new era of problems in scene understanding and it refers to the task of identifying an intersection from varied viewpoints in varied weather and lighting conditions. We focused on this significantly varied task approach to broaden the scope and generalizing capability of the results we compute. Both image annotation and intersection recognition pose some common challenges such as occlusion, perspective variations, distortions etc. While focusing on the image annotation task we further narrowed our domain by focusing on graph based methods. We again chose two different paradigms: a multiple kernel learning based non-deep learning approach and a deep-learning based approach, with a focus on bringing out contrast again. Through quantitative and qualitative results we show slightly boosted performance from the above mentioned paradigms. The intersection recognition task is relatively new in the field. Most of the work in field focuses on Places Recognition which utilized only single images. We focus on temporal information i.e. the traversal of the intersection/places as seen from a camera mounted on a vehicle. Through experiments we show a performance boost in intersection recognition from the inclusion of temporal information

Year of completion:  January 2020
 Advisor : Avinash Sharma

Related Publications

    Downloads

    thesis

    Computational Imaging Techniques to Recover Omni3D Structure & Surface Properties

    Rajat Aggarwal

    Abstract

    The process of imaging converts a 3D world into a 2D image. This process is inherently lossy, making the problem of understanding the world, ill-posed. During the imaging process, a light ray emanating from a source bounces of different surfaces, interacting with them according to the surface properties (albedo, color, specularity) before reaching the camera. The structure that we see in an image is primarily based on the final surface from which the light was reflected, except when that surface is highly specular (mirror-like). In this work, we explore imaging techniques the recover both the structure and surface properties of a scene. For structure recovery, we extend the idea of stereo imaging and present a practical solution to capture a complete 360◦ panorama using a single camera. Current approaches either use a moving camera for capturing multiple images of a scene, which are then stitched together to form the final panorama, or use multiple cameras that are synchronized. A moving camera limits the solution to static scenes, while multi-camera solutions require dedicated calibrated setups. Our approach improves upon the existing solutions in two significant ways: It solves the problem using a single camera, thus minimizing the calibration problem and providing us the ability to convert any digital camera into a stereo panoramic capture device. It captures all the light rays required for stereo panoramas in a single frame using a compact custom designed mirror, thus making the design practical to manufacture and easier to use. We analyze the optimality of the design as well as present panoramic stereo and depth estimation results. The methods for structure recovery, including stereo are often fooled when the surface is highly specular. To alleviate this, we propose an active-illumination based method to detect and segment mirror-like surfaces in a scene. In computer vision, many active illumination techniques employ Projector-Camera systems to extract useful information from the scenes. Known illumination patterns are projected onto the scene and their deformations in the captured images are then analyzed. We observe that the local frequencies in the captured pattern for the mirror-like surfaces is different from the projected pattern. This property allows us to design a custom Projector-Camera system to segment mirror-like surfaces by analyzing the local frequencies in the captured images. The system projects a sinusoidal pattern and capture the images from projector’s point of view. We present segmentation results for the scenes including multiple reflections and inter-reflections from the mirror-like surfaces. The method can further be used in the separation of direct and global components for the mirror-like surfaces by illuminating the non-mirror-like objects separately. We show how our method is also useful for accurate estimation of shape of the non-mirror-like regions in the presence of mirror-like regions in a scene.

    Year of completion:  November 2019
     Advisor : Anoop M Namboodiri

    Related Publications

      Downloads

      thesis

      Driver Attention Monitoring using Facial Features

      Isha Dua

      Abstract

      How can we assess the quality of human driving using AI? Driver inattention is one of the leading causes of vehicle crashes and incidents worldwide. Driver inattention includes driver fatigue leading to drowsiness and driver distraction, say due to the use of cellphone or rubbernecking, all of which leads to a lack of situational awareness. Hitherto, techniques presented to monitor driver attention evaluated factors such as fatigue and distraction independently. However, to develop a robust driver attention monitoring system, all the factors affecting a driver’s attention needs to be analyzed holistically. In this thesis, we present two novel approaches for driver attention analysis on the road using driver video and fusion of driver and road video.
      In the first approach, we propose the driver attention rating system that leverages the front camera of a windshield-mounted smartphone to monitor the driver attention by combining several features. We derive a driver attention rating by fusing spatio-temporal features based on the driver state and behavior such as head pose, eye gaze, eye closure, yawns, use of cellphones, etc. We present a few architec- tures for feature aggregation like AutoRate and Attention-based AutoRate. We perform an extensive evaluation of feature aggregation networks on real-world driving data and also data from controlled, static vehicle settings with 30 drivers in a large city. We compare the proposed method’s automatically- generated rating with the scores given by 5 human annotators. We introduce the kappa coefficient, an evaluation metric to compute the inter-rater agreement between the generated rating and the rating pro- vided by human annotators. We observe that Attention-based AutoRate outperforms other proposed designs for feature aggregation by 10%. Further, we use the learned temporal and spatial attention to visualize the key frame and the key action, which justifies the model’s predicted rating. Finally, to pro- vide driver-specific results, we fine-tune the Attention-based AutoRate model using the specific driver data to give personalized driver experience.

      Year of completion:  June 2020
       Advisor : Prof. C.V. Jawahar

      Related Publications

        Downloads

        thesis

        An investigation of the annotated data sparsity problem in the medical domain

        Pujitha Appan Kandala

        Abstract

        Diabetic retinopathy (DR) is the most common eye disease in people with diabetes. It affects them for significant number of years and can also lead to permanent blindness if left untreated. Early detection and treatment of DR is of utmost importance for the prevention of blindness. Hence, automatic disease detection and classification have been attracting much interest. High performance is critical in adoptionof such systems, which generally rely on training with a wide variety of annotated data. Availability of such varied annotated data in medical imaging is very scarce. The main focus of this thesis is to deal with the sparsity of annotated data and develop computer-aided diagnostic CAD systems which take less annotated data and yet give high accuracies. We propose three different solutions to address this problem. First, we propose a semi-supervised framework which paves way for including unlabeled data in training. A co-training framework is used in which features are extracted from a limited training set and independent models are learnt on each of the features, later the models are used to predict labels for new data. The highly confident labelled images from unlabelled set are added back to the training set and the process is continued, thus expanding the number of known labels. This framework is showcased on retinal neovascularization (NV) which is a critical stage of proliferative DR. The analysis of the results for detection of NV showed that an AUC of 0.985 with sensitivity of 96.2% at specificity of 92.6% which were superior to the existing models. Secondly, we propose crowdsourcing as a solution where we obtain annotations from a crowd and use them for training after refining. We employ a strategy to refine/overcome the noisy nature of crowdsourced annotations by i) assigning a reliability factor for each subject of the crowd based on their performance (at global and local levels) and experience and ii) requiring region of interest (ROI) markings rather than pixel-level markings from the crowd. We also show that these annotations are reliable by training a deep neural net (DNN) for detection of hard exudates which occur in mild non-proliferative DR. Experimental results obtained for hard exudate detection showed that training with refined crowdsourced data is effective as detection performance improves by 25% over training with just expert-markings. Lastly, we explore synthetic data generation as a solution to address this problem. We propose a novel method, based on generative adversarial networks (GAN), to generate images with lesions such that the overall severity level can be controlled. We showcase this approach for hard exudate and haemorrhage detection in retinal images with 4 levels of severity. These vary from mild to severe non-proliferativeDR. The synthetic data were also shown to be reliable for developing a CAD system for DR detection. Hard exudate/ haemorrhage detection was found to improve with inclusion of synthetic data in thetraining set with improvement in sensitivity of about 25% over training with just expert marked data.

         

        Year of completion:  November 2018
         Advisor : Jayanthi Sivaswamy

        Related Publications

          Downloads

          thesis

          Document Image Quality Assessment

          Pranjal Kumar Rai

          Abstract

          The amelioration in video capture technology has made recording videos very easy. The introduction of smaller affordable cameras which not only boast high-end specifications but are also capable of capturing videos at very high resolution (4K,8K and even 16K) have made recording of high quality videos accessible to everyone. Although this makes recording of videos very straightforward and effortless, a major part of the video production process - Editing is still labor intensive and requires skill and expertise. This thesis takes a step towards automating the editing process and making it less time consuming. In this thesis, (1) we explore a novel approach of automatically editing stage performances such that both the context of the scene as well as close-up details of the actors are shown. (2) We propose a new method to optimally re-target videos to any desired aspect ratio while retaining the salient regions that are derived using gaze tracking. Recordings of stage performances are easy to capture with a high-resolution camera, but are difficult to watch because the actors’ faces are too small. We present an approach to automatically create a split screen video that transforms these recordings to show both the context of the scene as well as close-up details of the actors. Given a static recording of a stage performance and the tracking information about the actors positions, our system generates videos showing a focus+context view based on computed close-up camera motions using crop-and zoom. The key to our approach is to compute these camera motions such that they are cinematically valid close-ups and to ensure that the set of views of the different actors are properly coordinated and presented. We pose the computation of camera motions as convex optimization that creates detailed views and smooth movements, subject to cinematic constraints such as not cutting faces with the edge of the frame. Additional constraints link the close up views of each actor, causing them to merge seamlessly when actors are close. Generated views are placed in a resulting layout that preserves the spatial relationships between actors. This eliminates the need for manual labour and expertise required for both capturing the performance and later editing it, instead the splitscreen of focus+context views allows the viewer to make an active decision on attending to whatever seems important. We also demonstrate our results on a variety of staged theater and dance performances. When videos are captured they are captured according to a specific aspect ratio keeping in mind the size of target screen in which they are meant to be viewed, this results in a inferior viewing experience when they are not watched on screens with their native aspect ratio. We present an approach to automatically retarget any given video to any desired aspect ratio while preserving its most salient regions obtained using gaze tracking. Our algorithm performs editing with cut, pan and zoom operations by optimizing the path of a cropping window within the original video while seeking to (i) preserve salient regions, and (ii) adhere to the principles of cinematography. The algorithm has two steps in total. The first step uses dynamic programming to find a cropping window path that maximizes gaze inclusion within the window and also tries to find the location of plausible new cuts (if required). The second step performs regularized convex optimization on the path obtained via dynamic programming to produce a smooth cropping window path comprised of piecewise linear, constant and parabolic segments. We test our re-editing algorithm on a diverse collection of movie and theater sequences. A study conducted with 16 users confirms that our retargeting algorithm results in a superior viewing experience as compared to gaze driven re-editing [30] and letterboxing methods, especially for wide-angle static camera recordings.

           

          Year of completion:  November 2018
           Advisor : Pranjal Kumar Rai

          Related Publications

            Downloads

            thesis

            More Articles …

            1. Computational Video Editing and Re-editing
            2. Geometric + Kinematic Priors and Part-based Graph Convolutional Network for Skeleton-based Human Action Recognition
            3. Towards Data-Driven Cinematography and Video Retargeting using Gaze
            4. Development and Tracking of Consensus Mesh for Monocular Depth Sequences