Abstract

Deep learning models are susceptible to backdoor attacks involving malicious attackers perturbing a small subset of training data with a trigger to causes misclassifications. Various triggers have been used including semantic triggers that are easily realizable without requiring attacker to manipulate the image. The emergence of generative AI has eased generation of varied poisoned samples. Robustness across types of triggers is crucial to effective defense. We propose Prototype Guided Backdoor Defense (PGBD), a robust post-hoc defense that scales across different trigger types, including previously unsolved semantic triggers. PGBD exploits displacements in the geometric spaces of activations to penalize movements towards the trigger. This is done using a novel sanitization loss of a post-hoc finetuning step. The geometric approach scales easily to all types of attacks. PGBD achieves better performance across all settings. We also present the first defense against a new semantic attack on celebrity face images.

Abstract

Creating photorealistic, animatable, and relightable 3D head avatars traditionally requires expensive Lightstage with multiple calibrated cameras, making it inaccessible for widespread adoption. To bridge this gap, we present a novel, cost-effective approach for creating high-quality relightable head avatars using only a smartphone equipped with polaroid filters. Our approach involves simultaneously capturing cross-polarized and parallel-polarized video streams in a dark room with a single point-light source, separating the skin's diffuse and specular components during dynamic facial performances. We introduce a hybrid representation that embeds 2D Gaussians in the UV space of a parametric head model, facilitating efficient real-time rendering while preserving high-fidelity geometric details. Our learning-based neural analysis-by-synthesis pipeline decouples pose and expression-dependent geometrical offsets from appearance, decomposing the surface into albedo, normal, and specular UV texture maps, along with the environment maps. We collect a unique dataset of various subjects performing diverse facial expressions and head movements.

Abstract

Linearly Transformed Spherical Distributions (LTSDs), a superset of the commonly known Linearly Transformed Cosines (LTCs) for analytic area light rendering, applied in the context of analytic area light rendering with participating media.

Abstract

Large-scale capturing of real-world scenes as 3D point clouds (e.g., using LIDAR scanning) generates billions of points that are challenging to visualize. High storage requirements prevent the quick and easy inspection of captured datasets on user-grade hardware. The fastest real-time rendering methods are limited by the available GPU memory and render only around 1 billion points interactively. We show that we can achieve state-of-the-art in both while simultaneously supporting datasets that surpass the capabilities of other methods. We present an on-the-fly point cloud decompression scheme that tightly integrates with software rasterization to reduce on-chip memory requirements by more than 4×. Our method compresses geometry losslessly and provides high visual quality at real-time framerates. We use a GPU-friendly, clipped Huffman encoding for compression. Point clouds are divided into equal-sized batches, which are Huffman-encoded independently. Batches are further subdivided to form easy-to-consume streams of data for massively parallel execution. The compressed point clouds are stored in an access-aware manner to achieve coherent GPU memory access

Abstract

The complex, glinty appearance of detailed normal-mapped surfaces at different scales requires expensive per-pixel Normal Distribution Function computations. Moreover, large light sources further compound this integration and increase the noise in the Monte Carlo renderer. Specialized rendering techniques that explicitly express the underlying normal distribution have been developed to improve performance for glinty surfaces controlled by a fixed material roughness. We present a new method that supports spatially varying roughness based on a neural histogram that computes per-pixel NDFs with arbitrary positions and sizes. Our representation is both memory and compute efficient. Additionally, we fully integrate direct illumination for all light directions in constant time. Our approach decouples roughness and normal distribution, allowing the live editing of the spatially varying roughness of complex normal-mapped objects. We demonstrate that our approach improves on previous work by achieving smaller footprints while offering GPU-friendly computation and compact representation.

Abstract

The focus of recent research has shifted from merely improving the metrics based performance of Deep Neural Networks (DNNs) to DNNs which are more interpretable to humans. The field of eXplainable Artificial Intelligence (XAI) has observed various techniques, including saliency-based and concept-based approaches. These approaches explain the model's decisions in simple human understandable terms called Concepts. Concepts are known to be the thinking ground of humans}. Explanations in terms of concepts enable detecting spurious correlations, inherent biases, or clever-hans. With the advent of concept-based explanations, a range of concept representation methods and automatic concept discovery algorithms have been introduced. Some recent works also use concepts for model improvement in terms of interpretability and generalization. We provide a systematic review and taxonomy of various concept representations and their discovery algorithms in DNNs, specifically in vision. We also provide details on concept-based model improvement literature marking the first comprehensive survey of these methods.

Abstract

We present a new additive image factorization technique that treats images to be composed of multiple latent specular components which can be simply estimated recursively by modulating the sparsity during decomposition. Our model-driven {\em RSFNet} estimates these factors by unrolling the optimization into network layers requiring only a few scalars to be learned. The resultant factors are interpretable by design and can be fused for different image enhancement tasks via a network or combined directly by the user in a controllable fashion. Based on RSFNet, we detail a zero-reference Low Light Enhancement (LLE) application trained without paired or unpaired supervision. Our system improves the state-of-the-art performance on standard benchmarks and achieves better generalization on multiple other datasets. We also integrate our factors with other task specific fusion networks for applications like deraining, deblurring and dehazing with negligible overhead thereby highlighting the multi-domain and multi-task generalizability of our proposed RSFNet. The code and data is released for reproducibility on the project homepage.

Abstract

Radiance Fields are being widely explored for 3D scene reconstruction and several downstream tasks, such as segmentation. Prior radiance field segmentation methods require scene-specific training to enable segmentation. We propose distilling semantic features into Radiance Fields in a generalisable fashion using GNT, a transformer-based architecture, enabling 3D reconstruction and multi-view segmentation on arbitrarily new scenes. By fine-tuning our method, any set of 2D features can be distilled into a radiance field, providing better multi-view consistency than the original features. We show multi-view segmentation results on standard datasets and compare our method against existing NeRF-based segmentation methods. We perform on par with the state-of-the-art scene-specific segmentation methods. Our approach and experiments bring generalisable NeRF methods one step closer to the contemporary NeRF literature.

Abstract

Radiance Fields (RF) are popular to represent casually-captured scenes for new view synthesis and several applications beyond it. Mixed reality on personal spaces needs understanding and manipulating scenes represented as RFs, with semantic segmentation of objects as an important step. Prior segmentation efforts show promise but don't scale to complex objects with diverse appearance. We present the ISRF method to interactively segment objects with fine structure and appearance. Nearest neighbor feature matching using distilled semantic features identifies high-confidence seed regions. Bilateral search in a joint spatio-semantic space grows the region to recover accurate segmentation. We show state-of-the-art results of segmenting objects from RFs and compositing them to another scene, changing appearance, etc., and an interactive segmentation tool that others can use.

Abstract

Humans use abstract concepts for understanding instead of hard features. Recent interpretability research has focused on human-centered concept explanations of neural networks. Concept Activation Vectors (CAVs) estimate a model’s sensitivity and possible biases to a given concept. In this paper, we extend CAVs from post- hoc analysis to ante-hoc training in order to reduce model bias through fine-tuning using an additional Concept Loss. Concepts were defined on the final layer of the network in the past. We generalize it to intermediate layers using class prototypes. This facilitates class learning in the last convolution layer, which is known to be most informative. We also introduce Concept Distillation to create richer concepts using a pre-trained knowledgeable model as the teacher. Our method can sensitize or desensitize a model towards concepts. We show applications of concept-sensitive training to debias several classification problems. We also use concepts to induce prior knowledge into IID, a reconstruction problem. Concept-sensitive training can improve model interpretability, reduce biases, and induce prior knowledge. Please visit https://avani17101.github.io/Concept-Distilllation/ for code and more details