Editing Neural Radiance Fields

Rahul Goel

Abstract

Neural Radiance Fields (NeRFs) have emerged as a pivotal advancement in computer graphics and vision. They provide a framework for rendering highly detailed novel view images from sparse multi- view input data. NeRFs use a continuous function to represent scenes that can be estimated using neural networks. This approach enables the generation of photorealistic images for static scenes. Outside the domain of image synthesis, NeRFs have been widely adopted as a representation of several downstream including but not limited to scene understanding, augemented reality, scene nav- igation, segmentation, and 3D asset generation. In this thesis, we explore upon the segmentation and editing capabilities in radiance fields. We propose a fast style transfer method that leverages multi-view consistent generation of stylized priors to change the appearance vectors in a Tensorial Radiance Field. Our method promises a speed-up of several orders of magnitude in applying style transfer and adheres to the colorscheme from the style image better than previous works. Next, we tackle the task of segmentation in radiance fields. Our method uses a grid-based feature field which allows extremely fast feature querying and searching. Combined with our stroke-based seg- mentation, this allws the user to interactively segment objects in a captured radiance field. We improve the state-of-the-art in terms of segmentation quality by a huge margin and in terms of segmentation time by orders of magnitude. Our method enables basic editing capabilities like translation, appearance editing, removal, and composition for which we show preliminary results. We further explore the problem of composition of radiance fields. Composition of two radiance fields using ray marching requires twice the amount of memory and compute. We use distillation to fuse multiple radiance fields into one to circumvent this problem. Our distillation process is roughly thrice as fast as re-training and produces a unified representation for radiance fields.

Year of completion:  April 2024
 Advisors : P J Narayanan

Related Publications

    Downloads

    thesis

    Neural Fields for Hand-object Interactions

    Chandradeep Pokhariya

    Abstract

    The hand is the most commonly used body part for interacting with our three-dimensional world. While it may seem ordinary, replicating hand movements with robots or in virtual/augmented reality is highly complex. Research on how hands interact with objects is crucial for advancing robotics, virtual reality, and human-computer interaction. Understanding hand movements and manipulation is critical to creating more intuitive and responsive technologies, which can significantly improve accuracy, efficiency, and scalability in various industries. Despite extensive research, programming robots to mimic human-hand interactions remains a challenging goal. One of the biggest challenges is collecting accurate 3D data for hand-object grasping. This process is complicated because of the hand’s flexibility and how hands and objects occlude in grasping poses. Collecting such data often requires expensive and sophisticated setups. However, recently, neural fields [1] have emerged, which can model 3D scenes using only multi-view images or videos. Neural fields use a continuous neural function to represent 3D scenes without needing 3D ground truth data, relying instead on differentiable rendering and multi-view photometric loss. With growing interest, these methods are becoming faster, more efficient, and better at modeling complex scenes. This thesis explores how neural fields can address two specific subproblems in hand-object interaction research. The first problem is generating novel grasps, which means predicting the final grasp pose of a hand based on its initial position and the object’s shape and location. The challenge is creating a generative model that can predict accurate grasp poses using only multi-view videos without 3D ground truth data. To solve this, we developed RealGrasper, a generative model that learns to predict grasp poses from multi-view data using photometric loss and other regularizations. The second problem is accurately capturing grasp poses and extracting contact points from multi-view videos. Current methods use the MANO model [2], which approximates hand shapes but lacks the details for precise contacts. Additionally, there is no easy way to get ground truth data for evaluating contact quality. To address this, we propose MANUS, a method for markerless grasp capture using articulated 3D Gaussians that reconstructs high-fidelity hand models from multi-view videos. We also created a large dataset, MANUS-Grasps, which includes multi-view videos of three subjects grasping over 30 objects. Furthermore, we developed a new way to capture and evaluate contacts, providing a contact metric for better assessment. We thoroughly evaluated our methods through detailed experiments, ablations, and comparisons, demonstrating that our approach outperforms existing state-of-the-art methods. We also summarize our contributions and discuss potential future directions in this field. We believe this thesis will help advance the research community further.

    Year of completion:  June 2024
     Advisors : Avinash Sharma,Srinath Sridhar

    Related Publications

      Downloads

      thesis

      Vulnerability of Neural Network based Speaker Recognition Systems

      Ritu Srivastava

      Abstract

      Speaker recognition (SR) involves automatic identification of individual speakers based on their voices, often representing acoustic traits as fixed-dimensional vectors through speaker embedding. A standard speaker recognition system (SRS) consists of three key phases: training, enrollment, and recognition. In each stage, acoustic features are extracted from raw speech signals using an acoustic feature extraction module, resulting in the acquisition of essential acoustic characteristics. Commonly used acoustic features include speech spectrogram, filter bank, and Mel-frequency cepstral coefficients. During the training stage, a background model is trained to establish a mapping from training voices to embeddings. The traditional background model employs a Gaussian Mixture Model (GMM) to generate identity-vector (ivector) embeddings. In contrast, more recent and promising background models leverage deep neural networks (DNNs) to generate deep embeddings, like xvector. In the enrollment stage, a voice spoken by an individual undergoing enrollment is mapped to an enrollment embedding using the previously trained background model. In the recognition stage, the process begins by retrieving the testing embedding of a given voice from the background model. Subsequently, the scoring module is engaged to measure the similarity between the enrollment and testing embeddings. The scoring module evaluates the similarity between the speaker and recorded embedding. Following the assessment, the scoring and decision module makes a decision based on the similarity score. A decision threshold is established, which serves as a criterion to determine whether the claimed identity of the speaker is accepted or rejected. The concept of voiceprint is rapidly gaining prominence as one of the emerging biometrics, primarily owing to its seamless integration with natural and human-centered Voice User Interface (VUI). The fast progress of Speaker Recognition Systems (SRSs) is intricately linked to the evolution of Neural Networks (NNs), with a particular emphasis on Deep Neural Networks (DNNs). With strides made in deep learning, Speaker Recognition (SR) has also benefitted and found extensive applications across hardware and software platforms. However, it has been shown that NNs are vulnerable to adversarial attacks, highlighting a challenge that needs to be addressed. Thus, even though users have the convenience of authentication with Speaker Recognition services, it has become evident that these solutions are vulnerable to adversarial attacks. This vulnerability highlights that Speaker Recognition (SR) is encountering security threats, raising significant concerns about user privacy. Adversarial attack was initially implemented with images, where an image classification model was successfully deceived using adversarial examples. Drawing inspiration from the progress made in adversarial attacks within the image domain, there is a growing interest in extending these techniques to the audio field. With emerging trends, convolutional neural networks have demonstrated instability to artificially crafted perturbations that remain undetectable to the human eye. Virtually every type of model, ranging from CNN to graphical neural network (GNN), has shown vulnerability to adversarial examples, particularly in the domain of image classification. Deep learning models typically get audio input by converting the audio into a spectrogram for further processing. A spectrogram serves as a condensed representation of an audio input. Given its image-like nature, the audio spectrogram is frequently used as input data for deep learning models, especially Convolutional Neural Networks (CNNs) adapted for audio tasks. CNN-based architectures were initially designed for image processing. This thesis contributes to the assessment of Convolutional Neural Networks (CNNs) for their resilience against adversarial attacks, a domain that is yet to be extensively investigated concerning endto-end trained CNNs for speaker recognition. This examination is essential for sustaining the integrity and security of speaker recognition systems. Our study fills this gap by exploring the variations of iterative Fast Gradient Sign Method (FGSM) to carry out adversarial attacks. We note that using a vanilla iterative FGSM technique can alter the identity of each speaker sample to any other speaker within the LibriSpeech dataset. Additionally, we introduce adversarial attacks specific to Mel spectrogram features by (a) constraining the number of manipulated pixels, (b) confining alterations to certain frequency bands, (c) limiting changes to particular time segments, and (d) employing a substitute model to generate the adversarial sample. Through comprehensive qualitative and quantitative analyses, we illustrate the vulnerability and counterintuitive behavior of existing CNN-based speaker recognition systems, wherein the predicted speaker identities can be inverted without discernible alterations in the audio. The samples are available at “https://advdemo.github.io/speech/".

      Year of completion:  June 2024
       Advisor : Vineet Gandhi

      Related Publications

        Downloads

        thesis

        Beyond Text: Expanding Speech Synthesis with Lip-to-Speech and Multi-Modal Fusion

        Neha Sahipjohn

        Abstract

        Speech constitutes a fundamental aspect of human communication. Therefore, the ability of computers to synthesize speech is paramount for achieving more natural human-computer interactions and increased accessibility, particularly for individuals with reading limitations. Recent advancements in AI and machine learning technologies, alongside generative AI techniques, have significantly improved speech synthesis quality. Text input serves as a common modality for speech synthesis, and Text-toSpeech (TTS) systems have achieved notable milestones in terms of intelligibility and naturalness. In this thesis, we propose a system to synthesize speech directly from lip movements and explore the idea of a unified speech synthesis model that can synthesize speech from different modalities, like text-only, video-only or combined text and video inputs. This facilitates applications in dubbing and accessibility initiatives aimed at providing voice to individuals who are unable to vocalize. This innovation promises streamlined communication in noisy environments as well. We propose a novel system for lip-to-speech synthesis that achieves state-of-the-art performance by leveraging advancements in selfsupervised learning and sequence-to-sequence networks. This enables the generation of highly intelligible and natural-sounding speech even with limited data. Existing lip-to-speech systems primarily focus on directly synthesizing speech or mel-spectrograms from lip movements. This often leads to compromised intelligibility and naturalness due to the entanglement of speech content with ambient information and speaker characteristics. We propose a modularized approach that uses representations that disentangle speech content from speaker characteristics, leading to superior performance. Our work sheds light on the information-rich nature of embedding spaces compared to tokenized representations. The system maps lip movement representations to disentangled speech representations, which are then fed into a vocoder for speech generation. Recognizing the potential applications in dubbing and the importance of synthesizing accurate speech, we explore a multimodal input setting by incorporating text alongside lip movements. Through extensive experimentation and evaluation across various datasets and metrics, we demonstrate the superior performance achieved by our proposed method. Our approach demonstrates high correctness and intelligibility, paving the way for practical deployment in real-world scenarios. Our work contributes significantly to advancing the field of lip-to-speech synthesis, offering a robust and versatile solution for generating natural-sounding speech from silent videos with broader implications for accessibility, human-computer interaction, and communication technology.

        Year of completion:  June 2024
         Advisor : Vineet Gandhi

        Related Publications

          Downloads

          thesis

          Unsupervised Learning of Disentangled Video Representation for Future Frame Prediction

          Ujjwal Tiwari

          Abstract

          Predicting what may happen in the future is a critical design element in developing an intelligent decision-making system. This thesis aims to shed some light on video prediction models that can predict future frames of a video sequence by observing a set of previously known frames. These models learn video representations encoding the causal rules that govern the physical world. Hence, these models have been extensively used in the design of various vision-guided robotic systems. These models also have applications in reinforcement learning, autonomous navigation, and healthcare. Video frame prediction remains challenging despite the availability of large amounts of video data and the recent progress of generative modeling techniques in synthesizing high-quality images. The challenges associated with predicting future frames can be attributed to two significant characteristics of video data - the high dimensionality of video frames and the stochastic nature of the motion exhibited in these video sequences. Existing video prediction models solve the challenge of predicting frames in high-dimensional pixel space by learning a low-dimensional disentangled video representation. These methods factorize video representations into dynamic and static components. The disentangled video representation is subsequently used for the downstream task of future frame prediction. In Chapter 3, we propose a mutual information-based predictive autoencoder, MIPAE, a self-supervised learning framework. The proposed framework factorizes the latent space representation of videos into two components - static content and a dynamic pose component. The MIPAE architecture comprises a content encoder, pose encoder, decoder, and a standard LSTM network. We train MIPAE using a twostep procedure, such that in the first step, the content encoder, pose encoder, and decoder are trained to learn disentangled frame representations. The content encoder is trained using the slow feature analysis constraint, while the pose encoder is trained using a novel mutual information loss term to achieve proper disentanglement. In the second step of our training methodology, we train an LSTM network to predict the low-dimensional pose representation of future frames. The predicted pose and learned content representations are decoded to generate future frames of a video sequence. In this thesis, we present detailed qualitative and quantitative results to compare the performance of our proposed MIPAE framework. We evaluate our approach on standard video prediction datasets like DSprites, MPI3D-real, and SMNIST using various visual quality assessment metrics, namely LPIPS, SSIM, and PSNR. We also present a metric based on mutual information gap, MIG, to quantitatively evaluate the degree of disentanglement between the factorized latent variables - pose and content. MIG score is subsequently used for a detailed comparative study of the proposed framework with other disentanglement-based video prediction approaches to showcase the efficacy of our disentanglement approach. We conclude our analysis by showcasing the visual superiority of the frames predicted by MIPAE. In Chapter 4, we explore the paradigm of stochastic video prediction models, which aim to capture the inherent uncertainty in real-world videos by using a stochastic latent variable to predict a different but plausible sequence of future frames corresponding to each sample of the stochastic latent variable. In our work, we modify the architecture of two stochastic video prediction models and apply a novel cycle consistency loss term to disentangle the video representation space into pose and content factors and model the uncertainty in the pose of various objects in the scene, to generate sharp and plausible frame predictions.

          Year of completion:  June 2024
           Advisor : Anoop M Namboodiri

          Related Publications

            Downloads

            thesis

            More Articles …

            1. Targeted Segmentation: Leveraging Localization with DAFT for Improved Medical Image Segmentation
            2. Estimating 3D Human Pose, Shape, and Correspondences from Monocular Input
            3. Quality Beyond Perception: Introducing Image Quality Metrics for Enhanced Facial and Fingerprint Recognition
            4. Towards Label Free Few Shot Learning : How Far Can We Go?